update

* Add initial template

* Second template

* feat: Add TextEmbeddingModule to AnyTextPipeline

* feat: Add AuxiliaryLatentModule template to AnyTextPipeline

* Add bert tokenizer from the anytext repo for now

* feat: Update AnyTextPipeline's modify_prompt method

This commit adds improvements to the modify_prompt method in the AnyTextPipeline class. The method now handles special characters and replaces selected string prompts with a placeholder. Additionally, it includes a check for Chinese text and translation using the trans_pipe.

* Fill in the `forward` pass of `AuxiliaryLatentModule`

* `make style && make quality`

* `chore: Update bert_tokenizer.py with a TODO comment suggesting the use of the transformers library`

* Update error handling to raise and logging

* Add `create_glyph_lines` function into `TextEmbeddingModule`

* make style

* Up

* Up

* Up

* Up

* Remove several comments

* refactor: Remove ControlNetConditioningEmbedding and update code accordingly

* Up

* Up

* up

* refactor: Update AnyTextPipeline to include new optional parameters

* up

* feat: Add OCR model and its components

* chore: Update `TextEmbeddingModule` to include OCR model components and dependencies

* chore: Update `AuxiliaryLatentModule` to include VAE model and its dependencies for masked image in the editing task

* `make style`

* refactor: Update `AnyTextPipeline`'s docstring

* Update `AuxiliaryLatentModule` to include info dictionary so that text processing is done once

* simplify

* `make style`

* Converting `TextEmbeddingModule` to ordinary `encode_prompt()` function

* Simplify for now

* `make style`

* Up

* feat: Add scripts to convert AnyText controlnet to diffusers

* `make style`

* Fix: Move glyph rendering to `TextEmbeddingModule` from `AuxiliaryLatentModule`

* make style

* Up

* Simplify

* Up

* feat: Add safetensors module for loading model file

* Fix device issues

* Up

* Up

* refactor: Simplify

* refactor: Simplify code for loading models and handling data types

* `make style`

* refactor: Update to() method in FrozenCLIPEmbedderT3 and TextEmbeddingModule

* refactor: Update dtype in embedding_manager.py to match proj.weight

* Up

* Add attribution and adaptation information to pipeline_anytext.py

* Update usage example

* Will refactor `controlnet_cond_embedding` initialization

* Add `AnyTextControlNetConditioningEmbedding` template

* Refactor organization

* style

* style

* Move custom blocks from `AuxiliaryLatentModule` to `AnyTextControlNetConditioningEmbedding`

* Follow one-file policy

* style

* [Docs] Update README and pipeline_anytext.py to use AnyTextControlNetModel

* [Docs] Update import statement for AnyTextControlNetModel in pipeline_anytext.py

* [Fix] Update import path for ControlNetModel, ControlNetOutput in anytext_controlnet.py

* Refactor AnyTextControlNet to use configurable conditioning embedding channels

* Complete control net conditioning embedding in AnyTextControlNetModel

* up

* [FIX] Ensure embeddings use correct device in AnyTextControlNetModel

* up

* up

* style

* [UPDATE] Revise README and example code for AnyTextPipeline integration with DiffusionPipeline

* [UPDATE] Update example code in anytext.py to use correct font file and improve clarity

* down

* [UPDATE] Refactor BasicTokenizer usage to a new Checker class for text processing

* update pillow

* [UPDATE] Remove commented-out code and unnecessary docstring in anytext.py and anytext_controlnet.py for improved clarity

* [REMOVE] Delete frozen_clip_embedder_t3.py as it is in the anytext.py file

* [UPDATE] Replace edict with dict for configuration in anytext.py and RecModel.py for consistency

* 🆙

* style

* [UPDATE] Revise README.md for clarity, remove unused imports in anytext.py, and add author credits in anytext_controlnet.py

* style

* Update examples/research_projects/anytext/README.md

Co-authored-by: Aryan <contact.aryanvs@gmail.com>

* Remove commented-out image preparation code in AnyTextPipeline

* Remove unnecessary blank line in README.md

.github/ISSUE_TEMPLATE/remote-vae-pilot-feedback.yml

@@ -0,0 +1,38 @@
+name: "\U0001F31F Remote VAE"
+description: Feedback for remote VAE pilot
+labels: [ "Remote VAE" ]
+
+body:
+  - type: textarea
+    id: positive
+    validations:
+      required: true
+    attributes:
+      label: Did you like the remote VAE solution?
+      description: |
+        If you liked it, we would appreciate it if you could elaborate what you liked.
+
+  - type: textarea
+    id: feedback
+    validations:
+      required: true
+    attributes:
+      label: What can be improved about the current solution?
+      description: |
+        Let us know the things you would like to see improved. Note that we will work optimizing the solution once the pilot is over and we have usage.
+
+  - type: textarea
+    id: others
+    validations:
+      required: true
+    attributes:
+      label: What other VAEs you would like to see if the pilot goes well?
+      description: |
+        Provide a list of the VAEs you would like to see in the future if the pilot goes well.
+
+  - type: textarea
+    id: additional-info
+    attributes:
+      label: Notify the members of the team
+      description: |
+        Tag the following folks when submitting this feedback: @hlky @sayakpaul

.github/workflows/nightly_tests.yml

@@ -418,6 +418,8 @@ jobs:
             test_location: "gguf"
           - backend: "torchao"
             test_location: "torchao"
+          - backend: "optimum_quanto"
+            test_location: "quanto"
     runs-on:
       group: aws-g6e-xlarge-plus
     container:

.github/workflows/pr_style_bot.yml

@@ -9,119 +9,43 @@ permissions:
   pull-requests: write
 
 jobs:
-  run-style-bot:
-    if: >
-      contains(github.event.comment.body, '@bot /style') &&
-      github.event.issue.pull_request != null
-    runs-on: ubuntu-latest
+  style:
+    uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@main
+    with:
+      python_quality_dependencies: "[quality]"
+      pre_commit_script_name: "Download and Compare files from the main branch"
+      pre_commit_script: |
+        echo "Downloading the files from the main branch"
 
-    steps:
-      - name: Extract PR details
-        id: pr_info
-        uses: actions/github-script@v6
-        with:
-          script: |
-            const prNumber = context.payload.issue.number;
-            const { data: pr } = await github.rest.pulls.get({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              pull_number: prNumber
-            });
-            
-            // We capture both the branch ref and the "full_name" of the head repo
-            // so that we can check out the correct repository & branch (including forks).
-            core.setOutput("prNumber", prNumber);
-            core.setOutput("headRef", pr.head.ref);
-            core.setOutput("headRepoFullName", pr.head.repo.full_name);
+        curl -o main_Makefile https://raw.githubusercontent.com/huggingface/diffusers/main/Makefile
+        curl -o main_setup.py https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/setup.py
+        curl -o main_check_doc_toc.py https://raw.githubusercontent.com/huggingface/diffusers/refs/heads/main/utils/check_doc_toc.py
 
-      - name: Check out PR branch
-        uses: actions/checkout@v3
-        env: 
-          HEADREPOFULLNAME: ${{ steps.pr_info.outputs.headRepoFullName }}
-          HEADREF: ${{ steps.pr_info.outputs.headRef }}
-        with:
-          # Instead of checking out the base repo, use the contributor's repo name
-          repository: ${{ env.HEADREPOFULLNAME }}
-          ref: ${{ env.HEADREF }}
-          # You may need fetch-depth: 0 for being able to push
-          fetch-depth: 0
-          token: ${{ secrets.GITHUB_TOKEN }}
-      
-      - name: Debug
-        env: 
-          HEADREPOFULLNAME: ${{ steps.pr_info.outputs.headRepoFullName }}
-          HEADREF: ${{ steps.pr_info.outputs.headRef }}
-          PRNUMBER: ${{ steps.pr_info.outputs.prNumber }}
-        run: |
-          echo "PR number: ${{ env.PRNUMBER }}"
-          echo "Head Ref: ${{ env.HEADREF }}"
-          echo "Head Repo Full Name: ${{ env.HEADREPOFULLNAME }}"
+        echo "Compare the files and raise error if needed"
 
-      - name: Set up Python
-        uses: actions/setup-python@v4
+        diff_failed=0
+        if ! diff -q main_Makefile Makefile; then
+          echo "Error: The Makefile has changed. Please ensure it matches the main branch."
+          diff_failed=1
+        fi
 
-      - name: Install dependencies
-        run: |
-          pip install .[quality]
+        if ! diff -q main_setup.py setup.py; then
+          echo "Error: The setup.py has changed. Please ensure it matches the main branch."
+          diff_failed=1
+        fi
 
-      - name: Download Makefile from main branch
-        run: |
-          curl -o main_Makefile https://raw.githubusercontent.com/huggingface/diffusers/main/Makefile
-        
-      - name: Compare Makefiles
-        run: |
-          if ! diff -q main_Makefile Makefile; then
-            echo "Error: The Makefile has changed. Please ensure it matches the main branch."
-            exit 1
-          fi
-          echo "No changes in Makefile. Proceeding..."
-          rm -rf main_Makefile
+        if ! diff -q main_check_doc_toc.py utils/check_doc_toc.py; then
+          echo "Error: The utils/check_doc_toc.py has changed. Please ensure it matches the main branch."
+          diff_failed=1
+        fi
 
-      - name: Run make style and make quality
-        run: |
-          make style && make quality
+        if [ $diff_failed -eq 1 ]; then
+          echo "❌ Error happened as we detected changes in the files that should not be changed ❌"
+          exit 1
+        fi
 
-      - name: Commit and push changes
-        id: commit_and_push
-        env: 
-          HEADREPOFULLNAME: ${{ steps.pr_info.outputs.headRepoFullName }}
-          HEADREF: ${{ steps.pr_info.outputs.headRef }}
-          PRNUMBER: ${{ steps.pr_info.outputs.prNumber }}
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          echo "HEADREPOFULLNAME: ${{ env.HEADREPOFULLNAME }}, HEADREF: ${{ env.HEADREF }}"
-          # Configure git with the Actions bot user
-          git config user.name "github-actions[bot]"
-          git config user.email "github-actions[bot]@users.noreply.github.com"
-
-          # Make sure your 'origin' remote is set to the contributor's fork
-          git remote set-url origin "https://x-access-token:${GITHUB_TOKEN}@github.com/${{ env.HEADREPOFULLNAME }}.git"
-
-          # If there are changes after running style/quality, commit them
-          if [ -n "$(git status --porcelain)" ]; then
-            git add .
-            git commit -m "Apply style fixes"
-            # Push to the original contributor's forked branch
-            git push origin HEAD:${{ env.HEADREF }}
-            echo "changes_pushed=true" >> $GITHUB_OUTPUT
-          else
-            echo "No changes to commit."
-            echo "changes_pushed=false" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Comment on PR with workflow run link
-        if: steps.commit_and_push.outputs.changes_pushed == 'true'
-        uses: actions/github-script@v6
-        with:
-          script: |
-            const prNumber = parseInt(process.env.prNumber, 10);
-            const runUrl = `${process.env.GITHUB_SERVER_URL}/${process.env.GITHUB_REPOSITORY}/actions/runs/${process.env.GITHUB_RUN_ID}`
-
-            await github.rest.issues.createComment({
-              owner: context.repo.owner,
-              repo: context.repo.repo,
-              issue_number: prNumber,
-              body: `Style fixes have been applied. [View the workflow run here](${runUrl}).`
-            });
-        env:
-          prNumber: ${{ steps.pr_info.outputs.prNumber }}
+        echo "No changes in the files. Proceeding..."
+        rm -rf main_Makefile main_setup.py main_check_doc_toc.py
+      style_command: "make style && make quality"
+    secrets:
+      bot_token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/pr_tests.yml

@@ -3,7 +3,6 @@ name: Fast tests for PRs
 on:
   pull_request:
     branches: [main]
-    types: [synchronize]
     paths:
       - "src/diffusers/**.py"
       - "benchmarks/**.py"

.github/workflows/pr_tests_gpu.yml

@@ -0,0 +1,250 @@
+name: Fast GPU Tests on PR 
+
+on:
+  pull_request:
+    branches: main
+    paths:
+      - "src/diffusers/models/modeling_utils.py"
+      - "src/diffusers/models/model_loading_utils.py"
+      - "src/diffusers/pipelines/pipeline_utils.py"
+      - "src/diffusers/pipeline_loading_utils.py"
+      - "src/diffusers/loaders/lora_base.py"
+      - "src/diffusers/loaders/lora_pipeline.py"
+      - "src/diffusers/loaders/peft.py"
+      - "tests/pipelines/test_pipelines_common.py"
+      - "tests/models/test_modeling_common.py"
+  workflow_dispatch:
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  DIFFUSERS_IS_CI: yes
+  OMP_NUM_THREADS: 8
+  MKL_NUM_THREADS: 8
+  HF_HUB_ENABLE_HF_TRANSFER: 1
+  PYTEST_TIMEOUT: 600
+  PIPELINE_USAGE_CUTOFF: 1000000000 # set high cutoff so that only always-test pipelines run
+
+jobs:
+  setup_torch_cuda_pipeline_matrix:
+    name: Setup Torch Pipelines CUDA Slow Tests Matrix
+    runs-on:
+      group: aws-general-8-plus
+    container:
+      image: diffusers/diffusers-pytorch-cpu
+    outputs:
+      pipeline_test_matrix: ${{ steps.fetch_pipeline_matrix.outputs.pipeline_test_matrix }}
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+      - name: Install dependencies
+        run: |
+          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+          python -m uv pip install -e [quality,test]
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: Fetch Pipeline Matrix
+        id: fetch_pipeline_matrix
+        run: |
+          matrix=$(python utils/fetch_torch_cuda_pipeline_test_matrix.py)
+          echo $matrix
+          echo "pipeline_test_matrix=$matrix" >> $GITHUB_OUTPUT
+      - name: Pipeline Tests Artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: test-pipelines.json
+          path: reports
+
+  torch_pipelines_cuda_tests:
+    name: Torch Pipelines CUDA Tests
+    needs: setup_torch_cuda_pipeline_matrix
+    strategy:
+      fail-fast: false
+      max-parallel: 8
+      matrix:
+        module: ${{ fromJson(needs.setup_torch_cuda_pipeline_matrix.outputs.pipeline_test_matrix) }}
+    runs-on:
+      group: aws-g4dn-2xlarge
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host --gpus 0
+    steps:
+      - name: Checkout diffusers
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 2
+
+      - name: NVIDIA-SMI
+        run: |
+          nvidia-smi
+      - name: Install dependencies
+        run: |
+          python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+          python -m uv pip install -e [quality,test]
+          pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+          pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+
+      - name: Environment
+        run: |
+          python utils/print_env.py
+      - name: Extract tests
+        id: extract_tests
+        run: |
+          pattern=$(python utils/extract_tests_from_mixin.py --type pipeline)
+          echo "$pattern" > /tmp/test_pattern.txt
+          echo "pattern_file=/tmp/test_pattern.txt" >> $GITHUB_OUTPUT
+
+      - name: PyTorch CUDA checkpoint tests on Ubuntu
+        env:
+          HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+          # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+          CUBLAS_WORKSPACE_CONFIG: :16:8
+        run: |
+          if [ "${{ matrix.module }}" = "ip_adapters" ]; then 
+              python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+              -s -v -k "not Flax and not Onnx" \
+              --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
+              tests/pipelines/${{ matrix.module }}
+          else 
+              pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
+              python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile \
+              -s -v -k "not Flax and not Onnx and $pattern" \
+              --make-reports=tests_pipeline_${{ matrix.module }}_cuda \
+              tests/pipelines/${{ matrix.module }}
+          fi 
+
+      - name: Failure short reports
+        if: ${{ failure() }}
+        run: |
+          cat reports/tests_pipeline_${{ matrix.module }}_cuda_stats.txt
+          cat reports/tests_pipeline_${{ matrix.module }}_cuda_failures_short.txt
+      - name: Test suite reports artifacts
+        if: ${{ always() }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: pipeline_${{ matrix.module }}_test_reports
+          path: reports
+
+  torch_cuda_tests:
+    name: Torch CUDA Tests
+    runs-on:
+      group: aws-g4dn-2xlarge
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --shm-size "16gb" --ipc host --gpus 0
+    defaults:
+      run:
+        shell: bash
+    strategy:
+      fail-fast: false
+      max-parallel: 2
+      matrix:
+        module: [models, schedulers, lora, others]
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test]
+        python -m uv pip install peft@git+https://github.com/huggingface/peft.git
+        pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git
+        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+
+    - name: Environment
+      run: |
+        python utils/print_env.py
+
+    - name: Extract tests
+      id: extract_tests
+      run: |
+        pattern=$(python utils/extract_tests_from_mixin.py --type ${{ matrix.module }})
+        echo "$pattern" > /tmp/test_pattern.txt
+        echo "pattern_file=/tmp/test_pattern.txt" >> $GITHUB_OUTPUT
+
+    - name: Run PyTorch CUDA tests
+      env:
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+        # https://pytorch.org/docs/stable/notes/randomness.html#avoiding-nondeterministic-algorithms
+        CUBLAS_WORKSPACE_CONFIG: :16:8
+      run: |
+        pattern=$(cat ${{ steps.extract_tests.outputs.pattern_file }})
+        if [ -z "$pattern" ]; then
+          python -m pytest -n 1 -sv --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx" tests/${{ matrix.module }} \
+          --make-reports=tests_torch_cuda_${{ matrix.module }}  
+        else
+          python -m pytest -n 1 -sv --max-worker-restart=0 --dist=loadfile -k "not Flax and not Onnx and $pattern" tests/${{ matrix.module }} \
+          --make-reports=tests_torch_cuda_${{ matrix.module }}  
+        fi
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/tests_torch_cuda_${{ matrix.module }}_stats.txt
+        cat reports/tests_torch_cuda_${{ matrix.module }}_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        name: torch_cuda_test_reports_${{ matrix.module }}
+        path: reports
+
+  run_examples_tests:
+    name: Examples PyTorch CUDA tests on Ubuntu
+        pip uninstall transformers -y && python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
+    runs-on:
+      group: aws-g4dn-2xlarge
+
+    container:
+      image: diffusers/diffusers-pytorch-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: |
+        nvidia-smi
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test,training]
+
+    - name: Environment
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python utils/print_env.py
+
+    - name: Run example tests on GPU
+      env:
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install timm
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v --make-reports=examples_torch_cuda examples/
+
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: |
+        cat reports/examples_torch_cuda_stats.txt
+        cat reports/examples_torch_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        name: examples_test_reports
+        path: reports
+

.github/workflows/push_tests.yml

@@ -1,13 +1,6 @@
 name: Fast GPU Tests on main
 
 on:
-  pull_request:
-    branches: main
-    paths:
-      - "src/diffusers/models/modeling_utils.py"
-      - "src/diffusers/models/model_loading_utils.py"
-      - "src/diffusers/pipelines/pipeline_utils.py"
-      - "src/diffusers/pipeline_loading_utils.py"
   workflow_dispatch:
   push:
     branches:
@@ -167,7 +160,6 @@ jobs:
         path: reports
 
   flax_tpu_tests:
-    if: ${{ github.event_name != 'pull_request' }}
     name: Flax TPU Tests
     runs-on:
       group: gcp-ct5lp-hightpu-8t
@@ -216,7 +208,6 @@ jobs:
         path: reports
 
   onnx_cuda_tests:
-    if: ${{ github.event_name != 'pull_request' }}
     name: ONNX CUDA Tests
     runs-on:
       group: aws-g4dn-2xlarge
@@ -265,7 +256,6 @@ jobs:
         path: reports
 
   run_torch_compile_tests:
-    if: ${{ github.event_name != 'pull_request' }}
     name: PyTorch Compile CUDA tests
 
     runs-on:
@@ -309,7 +299,6 @@ jobs:
         path: reports
 
   run_xformers_tests:
-    if: ${{ github.event_name != 'pull_request' }}
     name: PyTorch xformers CUDA tests
 
     runs-on:

docs/source/en/_toctree.yml

@@ -76,6 +76,14 @@
   - local: advanced_inference/outpaint
     title: Outpainting
   title: Advanced inference
+- sections:
+  - local: hybrid_inference/overview
+    title: Overview
+  - local: hybrid_inference/vae_decode
+    title: VAE Decode
+  - local: hybrid_inference/api_reference
+    title: API Reference
+  title: Hybrid Inference
 - sections:
   - local: using-diffusers/cogvideox
     title: CogVideoX
@@ -165,6 +173,8 @@
     title: gguf
   - local: quantization/torchao
     title: torchao
+  - local: quantization/quanto 
+    title: quanto
   title: Quantization Methods
 - sections:
   - local: optimization/fp16
@@ -282,6 +292,8 @@
         title: CogView4Transformer2DModel
       - local: api/models/dit_transformer2d
         title: DiTTransformer2DModel
+      - local: api/models/easyanimate_transformer3d
+        title: EasyAnimateTransformer3DModel
       - local: api/models/flux_transformer
         title: FluxTransformer2DModel
       - local: api/models/hunyuan_transformer2d
@@ -314,6 +326,8 @@
         title: Transformer2DModel
       - local: api/models/transformer_temporal
         title: TransformerTemporalModel
+      - local: api/models/wan_transformer_3d
+        title: WanTransformer3DModel
       title: Transformers
     - sections:
       - local: api/models/stable_cascade_unet
@@ -342,8 +356,12 @@
         title: AutoencoderKLHunyuanVideo
       - local: api/models/autoencoderkl_ltx_video
         title: AutoencoderKLLTXVideo
+      - local: api/models/autoencoderkl_magvit
+        title: AutoencoderKLMagvit
       - local: api/models/autoencoderkl_mochi
         title: AutoencoderKLMochi
+      - local: api/models/autoencoder_kl_wan
+        title: AutoencoderKLWan
       - local: api/models/asymmetricautoencoderkl
         title: AsymmetricAutoencoderKL
       - local: api/models/autoencoder_dc
@@ -418,6 +436,8 @@
       title: DiffEdit
     - local: api/pipelines/dit
       title: DiT
+    - local: api/pipelines/easyanimate
+      title: EasyAnimate
     - local: api/pipelines/flux
       title: Flux
     - local: api/pipelines/control_flux_inpaint
@@ -534,6 +554,8 @@
       title: UniDiffuser
     - local: api/pipelines/value_guided_sampling
       title: Value-guided sampling
+    - local: api/pipelines/wan
+      title: Wan
     - local: api/pipelines/wuerstchen
       title: Wuerstchen
     title: Pipelines
@@ -543,6 +565,10 @@
       title: Overview
     - local: api/schedulers/cm_stochastic_iterative
       title: CMStochasticIterativeScheduler
+    - local: api/schedulers/ddim_cogvideox
+      title: CogVideoXDDIMScheduler
+    - local: api/schedulers/multistep_dpm_solver_cogvideox
+      title: CogVideoXDPMScheduler
     - local: api/schedulers/consistency_decoder
       title: ConsistencyDecoderScheduler
     - local: api/schedulers/cosine_dpm

docs/source/en/api/models/autoencoder_kl_wan.md

@@ -0,0 +1,32 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLWan
+
+The 3D variational autoencoder (VAE) model with KL loss used in [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLWan
+
+vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
+```
+
+## AutoencoderKLWan
+
+[[autodoc]] AutoencoderKLWan
+  - decode
+  - all
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/autoencoderkl_magvit.md

@@ -0,0 +1,37 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# AutoencoderKLMagvit
+
+The 3D variational autoencoder (VAE) model with KL loss used in [EasyAnimate](https://github.com/aigc-apps/EasyAnimate) was introduced by Alibaba PAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import AutoencoderKLMagvit
+
+vae = AutoencoderKLMagvit.from_pretrained("alibaba-pai/EasyAnimateV5.1-12b-zh", subfolder="vae", torch_dtype=torch.float16).to("cuda")
+```
+
+## AutoencoderKLMagvit
+
+[[autodoc]] AutoencoderKLMagvit
+    - decode
+    - encode
+    - all
+
+## AutoencoderKLOutput
+
+[[autodoc]] models.autoencoders.autoencoder_kl.AutoencoderKLOutput
+
+## DecoderOutput
+
+[[autodoc]] models.autoencoders.vae.DecoderOutput

docs/source/en/api/models/easyanimate_transformer3d.md

@@ -0,0 +1,30 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# EasyAnimateTransformer3DModel
+
+A Diffusion Transformer model for 3D data from [EasyAnimate](https://github.com/aigc-apps/EasyAnimate) was introduced by Alibaba PAI.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import EasyAnimateTransformer3DModel
+
+transformer = EasyAnimateTransformer3DModel.from_pretrained("alibaba-pai/EasyAnimateV5.1-12b-zh", subfolder="transformer", torch_dtype=torch.float16).to("cuda")
+```
+
+## EasyAnimateTransformer3DModel
+
+[[autodoc]] EasyAnimateTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/models/wan_transformer_3d.md

@@ -0,0 +1,30 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License. -->
+
+# WanTransformer3DModel
+
+A Diffusion Transformer model for 3D video-like data was introduced in [Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+The model can be loaded with the following code snippet.
+
+```python
+from diffusers import WanTransformer3DModel
+
+transformer = WanTransformer3DModel.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="transformer", torch_dtype=torch.bfloat16)
+```
+
+## WanTransformer3DModel
+
+[[autodoc]] WanTransformer3DModel
+
+## Transformer2DModelOutput
+
+[[autodoc]] models.modeling_outputs.Transformer2DModelOutput

docs/source/en/api/pipelines/animatediff.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Text-to-Video Generation with AnimateDiff
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ## Overview
 
 [AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning](https://arxiv.org/abs/2307.04725) by Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, Bo Dai.

docs/source/en/api/pipelines/cogvideox.md

@@ -15,6 +15,10 @@
 
 # CogVideoX
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer](https://arxiv.org/abs/2408.06072) from Tsinghua University & ZhipuAI, by Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, Jie Tang.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/consisid.md

@@ -15,6 +15,10 @@
 
 # ConsisID
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [Identity-Preserving Text-to-Video Generation by Frequency Decomposition](https://arxiv.org/abs/2411.17440) from Peking University & University of Rochester & etc, by Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyang Ge, Yujun Shi, Liuhan Chen, Jiebo Luo, Li Yuan.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/control_flux_inpaint.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # FluxControlInpaint
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 FluxControlInpaintPipeline is an implementation of Inpainting for Flux.1 Depth/Canny models. It is a pipeline that allows you to inpaint images using the Flux.1 Depth/Canny models. The pipeline takes an image and a mask as input and returns the inpainted image.
 
 FLUX.1 Depth and Canny [dev] is a 12 billion parameter rectified flow transformer capable of generating an image based on a text description while following the structure of a given input image. **This is not a ControlNet model**.

docs/source/en/api/pipelines/controlnet.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # ControlNet
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
 
 With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.

docs/source/en/api/pipelines/controlnet_flux.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # ControlNet with Flux.1
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 FluxControlNetPipeline is an implementation of ControlNet for Flux.1.
 
 ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.

docs/source/en/api/pipelines/controlnet_sd3.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # ControlNet with Stable Diffusion 3
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 StableDiffusion3ControlNetPipeline is an implementation of ControlNet for Stable Diffusion 3.
 
 ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.

docs/source/en/api/pipelines/controlnet_sdxl.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # ControlNet with Stable Diffusion XL
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
 
 With a ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.

docs/source/en/api/pipelines/controlnet_union.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # ControlNetUnion
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ControlNetUnionModel is an implementation of ControlNet for Stable Diffusion XL.
 
 The ControlNet model was introduced in [ControlNetPlus](https://github.com/xinsir6/ControlNetPlus) by xinsir6. It supports multiple conditioning inputs without increasing computation.

docs/source/en/api/pipelines/controlnetxs.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # ControlNet-XS
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ControlNet-XS was introduced in [ControlNet-XS](https://vislearn.github.io/ControlNet-XS/) by Denis Zavadski and Carsten Rother. It is based on the observation that the control model in the [original ControlNet](https://huggingface.co/papers/2302.05543) can be made much smaller and still produce good results.
 
 Like the original ControlNet model, you can provide an additional control image to condition and control Stable Diffusion generation. For example, if you provide a depth map, the ControlNet model generates an image that'll preserve the spatial information from the depth map. It is a more flexible and accurate way to control the image generation process.

docs/source/en/api/pipelines/deepfloyd_if.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # DeepFloyd IF
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ## Overview
 
 DeepFloyd IF is a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding.

docs/source/en/api/pipelines/easyanimate.md

@@ -0,0 +1,88 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+-->
+
+# EasyAnimate
+[EasyAnimate](https://github.com/aigc-apps/EasyAnimate) by Alibaba PAI.
+
+The description from it's GitHub page:
+*EasyAnimate is a pipeline based on the transformer architecture, designed for generating AI images and videos, and for training baseline models and Lora models for Diffusion Transformer. We support direct prediction from pre-trained EasyAnimate models, allowing for the generation of videos with various resolutions, approximately 6 seconds in length, at 8fps (EasyAnimateV5.1, 1 to 49 frames). Additionally, users can train their own baseline and Lora models for specific style transformations.*
+
+This pipeline was contributed by [bubbliiiing](https://github.com/bubbliiiing). The original codebase can be found [here](https://huggingface.co/alibaba-pai). The original weights can be found under [hf.co/alibaba-pai](https://huggingface.co/alibaba-pai).
+
+There are two official EasyAnimate checkpoints for text-to-video and video-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh) | torch.float16 |
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-InP`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-InP) | torch.float16 |
+
+There is one official EasyAnimate checkpoints available for image-to-video and video-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-InP`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-InP) | torch.float16 |
+
+There are two official EasyAnimate checkpoints available for control-to-video.
+
+| checkpoints | recommended inference dtype |
+|:---:|:---:|
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-Control`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-Control) | torch.float16 |
+| [`alibaba-pai/EasyAnimateV5.1-12b-zh-Control-Camera`](https://huggingface.co/alibaba-pai/EasyAnimateV5.1-12b-zh-Control-Camera) | torch.float16 |
+
+For the EasyAnimateV5.1 series:
+- Text-to-video (T2V) and Image-to-video (I2V) works for multiple resolutions. The width and height can vary from 256 to 1024.
+- Both T2V and I2V models support generation with 1~49 frames and work best at this value. Exporting videos at 8 FPS is recommended.
+
+## Quantization
+
+Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
+
+Refer to the [Quantization](../../quantization/overview) overview to learn more about supported quantization backends and selecting a quantization backend that supports your use case. The example below demonstrates how to load a quantized [`EasyAnimatePipeline`] for inference with bitsandbytes.
+
+```py
+import torch
+from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, EasyAnimateTransformer3DModel, EasyAnimatePipeline
+from diffusers.utils import export_to_video
+
+quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
+transformer_8bit = EasyAnimateTransformer3DModel.from_pretrained(
+    "alibaba-pai/EasyAnimateV5.1-12b-zh",
+    subfolder="transformer",
+    quantization_config=quant_config,
+    torch_dtype=torch.float16,
+)
+
+pipeline = EasyAnimatePipeline.from_pretrained(
+    "alibaba-pai/EasyAnimateV5.1-12b-zh",
+    transformer=transformer_8bit,
+    torch_dtype=torch.float16,
+    device_map="balanced",
+)
+
+prompt = "A cat walks on the grass, realistic style."
+negative_prompt = "bad detailed"
+video = pipeline(prompt=prompt, negative_prompt=negative_prompt, num_frames=49, num_inference_steps=30).frames[0]
+export_to_video(video, "cat.mp4", fps=8)
+```
+
+## EasyAnimatePipeline
+
+[[autodoc]] EasyAnimatePipeline
+  - all
+  - __call__
+
+## EasyAnimatePipelineOutput
+
+[[autodoc]] pipelines.easyanimate.pipeline_output.EasyAnimatePipelineOutput

docs/source/en/api/pipelines/flux.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Flux
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Flux is a series of text-to-image generation models based on diffusion transformers. To know more about Flux, check out the original [blog post](https://blackforestlabs.ai/announcing-black-forest-labs/) by the creators of Flux, Black Forest Labs.
 
 Original model checkpoints for Flux can be found [here](https://huggingface.co/black-forest-labs). Original inference code can be found [here](https://github.com/black-forest-labs/flux).
@@ -355,8 +359,74 @@ image.save('flux_ip_adapter_output.jpg')
     <figcaption class="mt-2 text-sm text-center text-gray-500">IP-Adapter examples with prompt "wearing sunglasses"</figcaption>
 </div>
 
+## Optimize
 
-## Running FP16 inference
+Flux is a very large model and requires ~50GB of RAM/VRAM to load all the modeling components. Enable some of the optimizations below to lower the memory requirements.
+
+### Group offloading
+
+[Group offloading](../../optimization/memory#group-offloading) lowers VRAM usage by offloading groups of internal layers rather than the whole model or weights. You need to use [`~hooks.apply_group_offloading`] on all the model components of a pipeline. The `offload_type` parameter allows you to toggle between block and leaf-level offloading. Setting it to `leaf_level` offloads the lowest leaf-level parameters to the CPU instead of offloading at the module-level.
+
+On CUDA devices that support asynchronous data streaming, set `use_stream=True` to overlap data transfer and computation to accelerate inference.
+
+> [!TIP]
+> It is possible to mix block and leaf-level offloading for different components in a pipeline.
+
+```py
+import torch
+from diffusers import FluxPipeline
+from diffusers.hooks import apply_group_offloading
+
+model_id = "black-forest-labs/FLUX.1-dev"
+dtype = torch.bfloat16
+pipe = FluxPipeline.from_pretrained(
+	model_id,
+	torch_dtype=dtype,
+)
+
+apply_group_offloading(
+    pipe.transformer,
+    offload_type="leaf_level",
+    offload_device=torch.device("cpu"),
+    onload_device=torch.device("cuda"),
+    use_stream=True,
+)
+apply_group_offloading(
+    pipe.text_encoder, 
+    offload_device=torch.device("cpu"),
+    onload_device=torch.device("cuda"),
+    offload_type="leaf_level",
+    use_stream=True,
+)
+apply_group_offloading(
+    pipe.text_encoder_2, 
+    offload_device=torch.device("cpu"),
+    onload_device=torch.device("cuda"),
+    offload_type="leaf_level",
+    use_stream=True,
+)
+apply_group_offloading(
+    pipe.vae, 
+    offload_device=torch.device("cpu"),
+    onload_device=torch.device("cuda"),
+    offload_type="leaf_level",
+    use_stream=True,
+)
+
+prompt="A cat wearing sunglasses and working as a lifeguard at pool."
+
+generator = torch.Generator().manual_seed(181201)
+image = pipe(
+    prompt,
+    width=576,
+    height=1024,
+    num_inference_steps=30,
+    generator=generator
+).images[0]
+image
+```
+
+### Running FP16 inference
 
 Flux can generate high-quality images with FP16 (i.e. to accelerate inference on Turing/Volta GPUs) but produces different outputs compared to FP32/BF16. The issue is that some activations in the text encoders have to be clipped when running in FP16, which affects the overall image. Forcing text encoders to run with FP32 inference thus removes this output difference. See [here](https://github.com/huggingface/diffusers/pull/9097#issuecomment-2272292516) for details.
 
@@ -385,7 +455,7 @@ out = pipe(
 out.save("image.png")
 ```
 
-## Quantization
+### Quantization
 
 Quantization helps reduce the memory requirements of very large models by storing model weights in a lower precision data type. However, quantization may have varying impact on video quality depending on the video model.
 

docs/source/en/api/pipelines/hunyuan_video.md

@@ -14,6 +14,10 @@
 
 # HunyuanVideo
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [HunyuanVideo](https://www.arxiv.org/abs/2412.03603) by Tencent.
 
 *Recent advancements in video generation have significantly impacted daily life for both individuals and industries. However, the leading video generation models remain closed-source, resulting in a notable performance gap between industry capabilities and those available to the public. In this report, we introduce HunyuanVideo, an innovative open-source video foundation model that demonstrates performance in video generation comparable to, or even surpassing, that of leading closed-source models. HunyuanVideo encompasses a comprehensive framework that integrates several key elements, including data curation, advanced architectural design, progressive model scaling and training, and an efficient infrastructure tailored for large-scale model training and inference. As a result, we successfully trained a video generative model with over 13 billion parameters, making it the largest among all open-source models. We conducted extensive experiments and implemented a series of targeted designs to ensure high visual quality, motion dynamics, text-video alignment, and advanced filming techniques. According to evaluations by professionals, HunyuanVideo outperforms previous state-of-the-art models, including Runway Gen-3, Luma 1.6, and three top-performing Chinese video generative models. By releasing the code for the foundation model and its applications, we aim to bridge the gap between closed-source and open-source communities. This initiative will empower individuals within the community to experiment with their ideas, fostering a more dynamic and vibrant video generation ecosystem. The code is publicly available at [this https URL](https://github.com/tencent/HunyuanVideo).*
@@ -45,7 +49,8 @@ The following models are available for the image-to-video pipeline:
 
 | Model name | Description |
 |:---|:---|
-| [`https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
+| [`Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
+| [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |
 
 ## Quantization
 

docs/source/en/api/pipelines/kandinsky3.md

@@ -9,6 +9,10 @@ specific language governing permissions and limitations under the License.
 
 # Kandinsky 3
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Kandinsky 3 is created by [Vladimir Arkhipkin](https://github.com/oriBetelgeuse),[Anastasia Maltseva](https://github.com/NastyaMittseva),[Igor Pavlov](https://github.com/boomb0om),[Andrei Filatov](https://github.com/anvilarth),[Arseniy Shakhmatov](https://github.com/cene555),[Andrey Kuznetsov](https://github.com/kuznetsoffandrey),[Denis Dimitrov](https://github.com/denndimitrov), [Zein Shaheen](https://github.com/zeinsh)
 
 The description from it's GitHub page:

docs/source/en/api/pipelines/kolors.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Kolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ![](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/kolors/kolors_header_collage.png)
 
 Kolors is a large-scale text-to-image generation model based on latent diffusion, developed by [the Kuaishou Kolors team](https://github.com/Kwai-Kolors/Kolors). Trained on billions of text-image pairs, Kolors exhibits significant advantages over both open-source and closed-source models in visual quality, complex semantic accuracy, and text rendering for both Chinese and English characters. Furthermore, Kolors supports both Chinese and English inputs, demonstrating strong performance in understanding and generating Chinese-specific content. For more details, please refer to this [technical report](https://github.com/Kwai-Kolors/Kolors/blob/master/imgs/Kolors_paper.pdf).

docs/source/en/api/pipelines/latent_consistency_models.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Latent Consistency Models
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Latent Consistency Models (LCMs) were proposed in [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://huggingface.co/papers/2310.04378) by Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao.
 
 The abstract of the paper is as follows:

docs/source/en/api/pipelines/ledits_pp.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # LEDITS++
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 LEDITS++ was proposed in [LEDITS++: Limitless Image Editing using Text-to-Image Models](https://huggingface.co/papers/2311.16711) by Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, Apolinário Passos.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/ltx_video.md

@@ -14,6 +14,10 @@
 
 # LTX Video
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [LTX Video](https://huggingface.co/Lightricks/LTX-Video) is the first DiT-based video generation model capable of generating high-quality videos in real-time. It produces 24 FPS videos at a 768x512 resolution faster than they can be watched. Trained on a large-scale dataset of diverse videos, the model generates high-resolution videos with realistic and varied content. We provide a model for both text-to-video as well as image + text-to-video usecases.
 
 <Tip>

docs/source/en/api/pipelines/lumina2.md

@@ -14,6 +14,10 @@
 
 # Lumina2
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [Lumina Image 2.0: A Unified and Efficient Image Generative Model](https://huggingface.co/Alpha-VLLM/Lumina-Image-2.0) is a 2 billion parameter flow-based diffusion transformer capable of generating diverse images from text descriptions.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/marigold.md

@@ -1,4 +1,6 @@
-<!--Copyright 2024 Marigold authors and The HuggingFace Team. All rights reserved.
+<!--
+Copyright 2023-2025 Marigold Team, ETH Zürich. All rights reserved.
+Copyright 2024-2025 The HuggingFace Team. All rights reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
 the License. You may obtain a copy of the License at
@@ -10,67 +12,120 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Marigold Pipelines for Computer Vision Tasks
+# Marigold Computer Vision
 
 ![marigold](https://marigoldmonodepth.github.io/images/teaser_collage_compressed.jpg)
 
-Marigold was proposed in [Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation](https://huggingface.co/papers/2312.02145), a CVPR 2024 Oral paper by [Bingxin Ke](http://www.kebingxin.com/), [Anton Obukhov](https://www.obukhov.ai/), [Shengyu Huang](https://shengyuh.github.io/), [Nando Metzger](https://nandometzger.github.io/), [Rodrigo Caye Daudt](https://rcdaudt.github.io/), and [Konrad Schindler](https://scholar.google.com/citations?user=FZuNgqIAAAAJ&hl=en).
-The idea is to repurpose the rich generative prior of Text-to-Image Latent Diffusion Models (LDMs) for traditional computer vision tasks.
-Initially, this idea was explored to fine-tune Stable Diffusion for Monocular Depth Estimation, as shown in the teaser above.
-Later,
-- [Tianfu Wang](https://tianfwang.github.io/) trained the first Latent Consistency Model (LCM) of Marigold, which unlocked fast single-step inference;
-- [Kevin Qu](https://www.linkedin.com/in/kevin-qu-b3417621b/?locale=en_US) extended the approach to Surface Normals Estimation;
-- [Anton Obukhov](https://www.obukhov.ai/) contributed the pipelines and documentation into diffusers (enabled and supported by [YiYi Xu](https://yiyixuxu.github.io/) and [Sayak Paul](https://sayak.dev/)).
+Marigold was proposed in 
+[Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation](https://huggingface.co/papers/2312.02145), 
+a CVPR 2024 Oral paper by 
+[Bingxin Ke](http://www.kebingxin.com/), 
+[Anton Obukhov](https://www.obukhov.ai/), 
+[Shengyu Huang](https://shengyuh.github.io/), 
+[Nando Metzger](https://nandometzger.github.io/), 
+[Rodrigo Caye Daudt](https://rcdaudt.github.io/), and 
+[Konrad Schindler](https://scholar.google.com/citations?user=FZuNgqIAAAAJ&hl=en).
+The core idea is to **repurpose the generative prior of Text-to-Image Latent Diffusion Models (LDMs) for traditional 
+computer vision tasks**.
+This approach was explored by fine-tuning Stable Diffusion for **Monocular Depth Estimation**, as demonstrated in the 
+teaser above.
 
-The abstract from the paper is:
-
-*Monocular depth estimation is a fundamental computer vision task. Recovering 3D depth from a single image is geometrically ill-posed and requires scene understanding, so it is not surprising that the rise of deep learning has led to a breakthrough. The impressive progress of monocular depth estimators has mirrored the growth in model capacity, from relatively modest CNNs to large Transformer architectures. Still, monocular depth estimators tend to struggle when presented with images with unfamiliar content and layout, since their knowledge of the visual world is restricted by the data seen during training, and challenged by zero-shot generalization to new domains. This motivates us to explore whether the extensive priors captured in recent generative diffusion models can enable better, more generalizable depth estimation. We introduce Marigold, a method for affine-invariant monocular depth estimation that is derived from Stable Diffusion and retains its rich prior knowledge. The estimator can be fine-tuned in a couple of days on a single GPU using only synthetic training data. It delivers state-of-the-art performance across a wide range of datasets, including over 20% performance gains in specific cases. Project page: https://marigoldmonodepth.github.io.*
-
-## Available Pipelines
-
-Each pipeline supports one Computer Vision task, which takes an input RGB image as input and produces a *prediction* of the modality of interest, such as a depth map of the input image.
-Currently, the following tasks are implemented:
-
-| Pipeline                                                                                                                                    | Predicted Modalities                                                                                             |                                                                       Demos                                                                        |
-|---------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------|:--------------------------------------------------------------------------------------------------------------------------------------------------:|
-| [MarigoldDepthPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_depth.py)     | [Depth](https://en.wikipedia.org/wiki/Depth_map), [Disparity](https://en.wikipedia.org/wiki/Binocular_disparity) | [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-lcm), [Slow Original Demo (DDIM)](https://huggingface.co/spaces/prs-eth/marigold) |
-| [MarigoldNormalsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_normals.py) | [Surface normals](https://en.wikipedia.org/wiki/Normal_mapping)                                                  |                                   [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-normals-lcm)                                    |
-
-
-## Available Checkpoints
-
-The original checkpoints can be found under the [PRS-ETH](https://huggingface.co/prs-eth/) Hugging Face organization.
+Marigold was later extended in the follow-up paper, 
+[Marigold: Affordable Adaptation of Diffusion-Based Image Generators for Image Analysis](https://huggingface.co/papers/2312.02145), 
+authored by 
+[Bingxin Ke](http://www.kebingxin.com/), 
+[Kevin Qu](https://www.linkedin.com/in/kevin-qu-b3417621b/?locale=en_US), 
+[Tianfu Wang](https://tianfwang.github.io/), 
+[Nando Metzger](https://nandometzger.github.io/), 
+[Shengyu Huang](https://shengyuh.github.io/), 
+[Bo Li](https://www.linkedin.com/in/bobboli0202/), 
+[Anton Obukhov](https://www.obukhov.ai/), and 
+[Konrad Schindler](https://scholar.google.com/citations?user=FZuNgqIAAAAJ&hl=en).
+This work expanded Marigold to support new modalities such as **Surface Normals** and **Intrinsic Image Decomposition** 
+(IID), introduced a training protocol for **Latent Consistency Models** (LCM), and demonstrated **High-Resolution** (HR) 
+processing capability.
 
 <Tip>
 
-Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines. Also, to know more about reducing the memory usage of this pipeline, refer to the ["Reduce memory usage"] section [here](../../using-diffusers/svd#reduce-memory-usage).
+The early Marigold models (`v1-0` and earlier) were optimized for best results with at least 10 inference steps.
+LCM models were later developed to enable high-quality inference in just 1 to 4 steps.
+Marigold models `v1-1` and later use the DDIM scheduler to achieve optimal 
+results in as few as 1 to 4 steps.
+
+</Tip>
+
+## Available Pipelines
+
+Each pipeline is tailored for a specific computer vision task, processing an input RGB image and generating a 
+corresponding prediction.
+Currently, the following computer vision tasks are implemented:
+
+| Pipeline                                                                                                                                          | Recommended Model Checkpoints                                                                                                                                                                           |                              Spaces (Interactive Apps)                               | Predicted Modalities                                                                                                                                                               |
+|---------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------:|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [MarigoldDepthPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_depth.py)           | [prs-eth/marigold-depth-v1-1](https://huggingface.co/prs-eth/marigold-depth-v1-1)                                                                                                                       |          [Depth Estimation](https://huggingface.co/spaces/prs-eth/marigold)          | [Depth](https://en.wikipedia.org/wiki/Depth_map), [Disparity](https://en.wikipedia.org/wiki/Binocular_disparity)                                                                   |
+| [MarigoldNormalsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_normals.py)       | [prs-eth/marigold-normals-v1-1](https://huggingface.co/prs-eth/marigold-normals-v1-1)                                                                                                                   | [Surface Normals Estimation](https://huggingface.co/spaces/prs-eth/marigold-normals) | [Surface normals](https://en.wikipedia.org/wiki/Normal_mapping)                                                                                                                    |
+| [MarigoldIntrinsicsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_intrinsics.py) | [prs-eth/marigold-iid-appearance-v1-1](https://huggingface.co/prs-eth/marigold-iid-appearance-v1-1),<br>[prs-eth/marigold-iid-lighting-v1-1](https://huggingface.co/prs-eth/marigold-iid-lighting-v1-1) | [Intrinsic Image Decomposition](https://huggingface.co/spaces/prs-eth/marigold-iid)  | [Albedo](https://en.wikipedia.org/wiki/Albedo), [Materials](https://www.n.aiq3d.com/wiki/roughnessmetalnessao-map), [Lighting](https://en.wikipedia.org/wiki/Diffuse_reflection)   |
+
+## Available Checkpoints
+
+All original checkpoints are available under the [PRS-ETH](https://huggingface.co/prs-eth/) organization on Hugging Face.
+They are designed for use with diffusers pipelines and the [original codebase](https://github.com/prs-eth/marigold), which can also be used to train 
+new model checkpoints.
+The following is a summary of the recommended checkpoints, all of which produce reliable results with 1 to 4 steps. 
+
+| Checkpoint                                                                                          | Modality     | Comment                                                                                                                                                                              |
+|-----------------------------------------------------------------------------------------------------|--------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [prs-eth/marigold-depth-v1-1](https://huggingface.co/prs-eth/marigold-depth-v1-1)                   | Depth        | Affine-invariant depth prediction assigns each pixel a value between 0 (near plane) and 1 (far plane), with both planes determined by the model during inference.                    |
+| [prs-eth/marigold-normals-v0-1](https://huggingface.co/prs-eth/marigold-normals-v0-1)               | Normals      | The surface normals predictions are unit-length 3D vectors in the screen space camera, with values in the range from -1 to 1.                                                        |
+| [prs-eth/marigold-iid-appearance-v1-1](https://huggingface.co/prs-eth/marigold-iid-appearance-v1-1) | Intrinsics   | InteriorVerse decomposition is comprised of Albedo and two BRDF material properties: Roughness and Metallicity.                                                                      | 
+| [prs-eth/marigold-iid-lighting-v1-1](https://huggingface.co/prs-eth/marigold-iid-lighting-v1-1)     | Intrinsics   | HyperSim decomposition of an image &nbsp\\(I\\)&nbsp is comprised of Albedo &nbsp\\(A\\), Diffuse shading &nbsp\\(S\\), and Non-diffuse residual &nbsp\\(R\\): &nbsp\\(I = A*S+R\\). |
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff 
+between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to 
+efficiently load the same components into multiple pipelines. 
+Also, to know more about reducing the memory usage of this pipeline, refer to the ["Reduce memory usage"] section 
+[here](../../using-diffusers/svd#reduce-memory-usage).
 
 </Tip>
 
 <Tip warning={true}>
 
-Marigold pipelines were designed and tested only with `DDIMScheduler` and `LCMScheduler`.
-Depending on the scheduler, the number of inference steps required to get reliable predictions varies, and there is no universal value that works best across schedulers.
-Because of that, the default value of `num_inference_steps` in the `__call__` method of the pipeline is set to `None` (see the API reference).
-Unless set explicitly, its value will be taken from the checkpoint configuration `model_index.json`.
-This is done to ensure high-quality predictions when calling the pipeline with just the `image` argument.
+Marigold pipelines were designed and tested with the scheduler embedded in the model checkpoint.
+The optimal number of inference steps varies by scheduler, with no universal value that works best across all cases.
+To accommodate this, the `num_inference_steps` parameter in the pipeline's `__call__` method defaults to `None` (see the 
+API reference).
+Unless set explicitly, it inherits the value from the `default_denoising_steps` field in the checkpoint configuration 
+file (`model_index.json`).
+This ensures high-quality predictions when invoking the pipeline with only the `image` argument.
 
 </Tip>
 
-See also Marigold [usage examples](marigold_usage).
+See also Marigold [usage examples](../../using-diffusers/marigold_usage).
+
+## Marigold Depth Prediction API
 
-## MarigoldDepthPipeline
 [[autodoc]] MarigoldDepthPipeline
-	- all
 	- __call__
 
-## MarigoldNormalsPipeline
-[[autodoc]] MarigoldNormalsPipeline
-	- all
-	- __call__
-
-## MarigoldDepthOutput
 [[autodoc]] pipelines.marigold.pipeline_marigold_depth.MarigoldDepthOutput
 
-## MarigoldNormalsOutput
-[[autodoc]] pipelines.marigold.pipeline_marigold_normals.MarigoldNormalsOutput
+[[autodoc]] pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_depth
+
+## Marigold Normals Estimation API
+[[autodoc]] MarigoldNormalsPipeline
+	- __call__
+
+[[autodoc]] pipelines.marigold.pipeline_marigold_normals.MarigoldNormalsOutput
+
+[[autodoc]] pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_normals
+
+## Marigold Intrinsic Image Decomposition API
+
+[[autodoc]] MarigoldIntrinsicsPipeline
+	- __call__
+
+[[autodoc]] pipelines.marigold.pipeline_marigold_intrinsics.MarigoldIntrinsicsOutput
+
+[[autodoc]] pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_intrinsics

docs/source/en/api/pipelines/mochi.md

@@ -15,6 +15,10 @@
 
 # Mochi 1 Preview
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 > [!TIP]
 > Only a research preview of the model weights is available at the moment.
 

docs/source/en/api/pipelines/overview.md

@@ -54,7 +54,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [DiT](dit) | text2image |
 | [Flux](flux) | text2image |
 | [Hunyuan-DiT](hunyuandit) | text2image |
-| [I2VGen-XL](i2vgenxl) | text2video |
+| [I2VGen-XL](i2vgenxl) | image2video |
 | [InstructPix2Pix](pix2pix) | image editing |
 | [Kandinsky 2.1](kandinsky) | text2image, image2image, inpainting, interpolation |
 | [Kandinsky 2.2](kandinsky_v22) | text2image, image2image, inpainting |
@@ -65,7 +65,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [Latte](latte) | text2image |
 | [LEDITS++](ledits_pp) | image editing |
 | [Lumina-T2X](lumina) | text2image |
-| [Marigold](marigold) | depth |
+| [Marigold](marigold) | depth-estimation, normals-estimation, intrinsic-decomposition |
 | [MultiDiffusion](panorama) | text2image |
 | [MusicLDM](musicldm) | text2audio |
 | [PAG](pag) | text2image |

docs/source/en/api/pipelines/pag.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Perturbed-Attention Guidance
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [Perturbed-Attention Guidance (PAG)](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) is a new diffusion sampling guidance that improves sample quality across both unconditional and conditional settings, achieving this without requiring further training or the integration of external modules.
 
 PAG was introduced in [Self-Rectifying Diffusion Sampling with Perturbed-Attention Guidance](https://huggingface.co/papers/2403.17377) by Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin and Seungryong Kim.

docs/source/en/api/pipelines/panorama.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # MultiDiffusion
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation](https://huggingface.co/papers/2302.08113) is by Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/pia.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Image-to-Video Generation with PIA (Personalized Image Animator)
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 ## Overview
 
 [PIA: Your Personalized Image Animator via Plug-and-Play Modules in Text-to-Image Models](https://arxiv.org/abs/2312.13964) by Yiming Zhang, Zhening Xing, Yanhong Zeng, Youqing Fang, Kai Chen

docs/source/en/api/pipelines/pix2pix.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # InstructPix2Pix
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [InstructPix2Pix: Learning to Follow Image Editing Instructions](https://huggingface.co/papers/2211.09800) is by Tim Brooks, Aleksander Holynski and Alexei A. Efros.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/sana.md

@@ -14,6 +14,10 @@
 
 # SanaPipeline
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers](https://huggingface.co/papers/2410.10629) from NVIDIA and MIT HAN Lab, by Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, Song Han.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/stable_diffusion/depth2img.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Depth-to-image
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 The Stable Diffusion model can also infer depth based on an image using [MiDaS](https://github.com/isl-org/MiDaS). This allows you to pass a text prompt and an initial image to condition the generation of new images as well as a `depth_map` to preserve the image structure.
 
 <Tip>

docs/source/en/api/pipelines/stable_diffusion/img2img.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Image-to-image
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 The Stable Diffusion model can also be applied to image-to-image generation by passing a text prompt and an initial image to condition the generation of new images.
 
 The [`StableDiffusionImg2ImgPipeline`] uses the diffusion-denoising mechanism proposed in [SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations](https://huggingface.co/papers/2108.01073) by Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, Stefano Ermon.

docs/source/en/api/pipelines/stable_diffusion/inpaint.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Inpainting
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 The Stable Diffusion model can also be applied to inpainting which lets you edit specific parts of an image by providing a mask and a text prompt using Stable Diffusion.
 
 ## Tips

docs/source/en/api/pipelines/stable_diffusion/ldm3d_diffusion.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Text-to-(RGB, depth)
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 LDM3D was proposed in [LDM3D: Latent Diffusion Model for 3D](https://huggingface.co/papers/2305.10853) by Gabriela Ben Melech Stan, Diana Wofk, Scottie Fox, Alex Redden, Will Saxton, Jean Yu, Estelle Aflalo, Shao-Yen Tseng, Fabio Nonato, Matthias Muller, and Vasudev Lal. LDM3D generates an image and a depth map from a given text prompt unlike the existing text-to-image diffusion models such as [Stable Diffusion](./overview) which only generates an image. With almost the same number of parameters, LDM3D achieves to create a latent space that can compress both the RGB images and the depth maps.
 
 Two checkpoints are available for use:

docs/source/en/api/pipelines/stable_diffusion/overview.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Stable Diffusion pipelines
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Stable Diffusion is a text-to-image latent diffusion model created by the researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/) and [LAION](https://laion.ai/). Latent diffusion applies the diffusion process over a lower dimensional latent space to reduce memory and compute complexity. This specific type of diffusion model was proposed in [High-Resolution Image Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer.
 
 Stable Diffusion is trained on 512x512 images from a subset of the LAION-5B dataset. This model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts. With its 860M UNet and 123M text encoder, the model is relatively lightweight and can run on consumer GPUs.

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_3.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Stable Diffusion 3
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Stable Diffusion 3 (SD3) was proposed in [Scaling Rectified Flow Transformers for High-Resolution Image Synthesis](https://arxiv.org/pdf/2403.03206.pdf) by Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/stable_diffusion/stable_diffusion_xl.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Stable Diffusion XL
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Stable Diffusion XL (SDXL) was proposed in [SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis](https://huggingface.co/papers/2307.01952) by Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/stable_diffusion/text2img.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Text-to-image
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 The Stable Diffusion model was created by researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/), [Runway](https://github.com/runwayml), and [LAION](https://laion.ai/). The [`StableDiffusionPipeline`] is capable of generating photorealistic images given any text input. It's trained on 512x512 images from a subset of the LAION-5B dataset. This model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts. With its 860M UNet and 123M text encoder, the model is relatively lightweight and can run on consumer GPUs. Latent diffusion is the research on top of which Stable Diffusion was built. It was proposed in [High-Resolution Image Synthesis with Latent Diffusion Models](https://huggingface.co/papers/2112.10752) by Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/stable_diffusion/upscale.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Super-resolution
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 The Stable Diffusion upscaler diffusion model was created by the researchers and engineers from [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/), and [LAION](https://laion.ai/). It is used to enhance the resolution of input images by a factor of 4.
 
 <Tip>

docs/source/en/api/pipelines/stable_unclip.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Stable unCLIP
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 Stable unCLIP checkpoints are finetuned from [Stable Diffusion 2.1](./stable_diffusion/stable_diffusion_2) checkpoints to condition on CLIP image embeddings.
 Stable unCLIP still conditions on text embeddings. Given the two separate conditionings, stable unCLIP can be used
 for text guided image variation. When combined with an unCLIP prior, it can also be used for full text to image generation.

docs/source/en/api/pipelines/text_to_video.md

@@ -18,6 +18,10 @@ specific language governing permissions and limitations under the License.
 
 # Text-to-video
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [ModelScope Text-to-Video Technical Report](https://arxiv.org/abs/2308.06571) is by Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, Shiwei Zhang.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/text_to_video_zero.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Text2Video-Zero
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 [Text2Video-Zero: Text-to-Image Diffusion Models are Zero-Shot Video Generators](https://huggingface.co/papers/2303.13439) is by Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, [Zhangyang Wang](https://www.ece.utexas.edu/people/faculty/atlas-wang), Shant Navasardyan, [Humphrey Shi](https://www.humphreyshi.com).
 
 Text2Video-Zero enables zero-shot video generation using either:

docs/source/en/api/pipelines/unidiffuser.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # UniDiffuser
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 The UniDiffuser model was proposed in [One Transformer Fits All Distributions in Multi-Modal Diffusion at Scale](https://huggingface.co/papers/2303.06555) by Fan Bao, Shen Nie, Kaiwen Xue, Chongxuan Li, Shi Pu, Yaole Wang, Gang Yue, Yue Cao, Hang Su, Jun Zhu.
 
 The abstract from the paper is:

docs/source/en/api/pipelines/wan.md

@@ -0,0 +1,78 @@
+<!-- Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License. -->
+
+# Wan
+
+[Wan 2.1](https://github.com/Wan-Video/Wan2.1) by the Alibaba Wan Team.
+
+<!-- TODO(aryan): update abstract once paper is out -->
+
+<Tip>
+
+Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers) to learn how to explore the tradeoff between scheduler speed and quality, and see the [reuse components across pipelines](../../using-diffusers/loading#reuse-a-pipeline) section to learn how to efficiently load the same components into multiple pipelines.
+
+</Tip>
+
+Recommendations for inference:
+- VAE in `torch.float32` for better decoding quality.
+- `num_frames` should be of the form `4 * k + 1`, for example `49` or `81`.
+- For smaller resolution videos, try lower values of `shift` (between `2.0` to `5.0`) in the [Scheduler](https://huggingface.co/docs/diffusers/main/en/api/schedulers/flow_match_euler_discrete#diffusers.FlowMatchEulerDiscreteScheduler.shift). For larger resolution videos, try higher values (between `7.0` and `12.0`). The default value is `3.0` for Wan.
+
+### Using a custom scheduler
+
+Wan can be used with many different schedulers, each with their own benefits regarding speed and generation quality. By default, Wan uses the `UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=3.0)` scheduler. You can use a different scheduler as follows:
+
+```python
+from diffusers import FlowMatchEulerDiscreteScheduler, UniPCMultistepScheduler, WanPipeline
+
+scheduler_a = FlowMatchEulerDiscreteScheduler(shift=5.0)
+scheduler_b = UniPCMultistepScheduler(prediction_type="flow_prediction", use_flow_sigmas=True, flow_shift=4.0)
+
+pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", scheduler=<CUSTOM_SCHEDULER_HERE>)
+
+# or,
+pipe.scheduler = <CUSTOM_SCHEDULER_HERE>
+```
+
+### Using single file loading with Wan
+
+The `WanTransformer3DModel` and `AutoencoderKLWan` models support loading checkpoints in their original format via the `from_single_file` loading 
+method. 
+
+
+```python
+import torch
+from diffusers import WanPipeline, WanTransformer3DModel
+
+ckpt_path = "https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/diffusion_models/wan2.1_t2v_1.3B_bf16.safetensors"
+transformer = WanTransformer3DModel.from_single_file(ckpt_path, torch_dtype=torch.bfloat16)
+
+pipe = WanPipeline.from_pretrained("Wan-AI/Wan2.1-T2V-1.3B-Diffusers", transformer=transformer)
+```
+
+## WanPipeline
+
+[[autodoc]] WanPipeline
+  - all
+  - __call__
+
+## WanImageToVideoPipeline
+
+[[autodoc]] WanImageToVideoPipeline
+  - all
+  - __call__
+
+## WanPipelineOutput
+
+[[autodoc]] pipelines.wan.pipeline_output.WanPipelineOutput

docs/source/en/api/pipelines/wuerstchen.md

@@ -12,6 +12,10 @@ specific language governing permissions and limitations under the License.
 
 # Würstchen
 
+<div class="flex flex-wrap space-x-1">
+  <img alt="LoRA" src="https://img.shields.io/badge/LoRA-d8b4fe?style=flat"/>
+</div>
+
 <img src="https://github.com/dome272/Wuerstchen/assets/61938694/0617c863-165a-43ee-9303-2a17299a0cf9">
 
 [Wuerstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models](https://huggingface.co/papers/2306.00637) is by Pablo Pernias, Dominic Rampas, Mats L. Richter and Christopher Pal and Marc Aubreville.

docs/source/en/api/quantization.md

@@ -31,6 +31,11 @@ Learn how to quantize models in the [Quantization](../quantization/overview) gui
 ## GGUFQuantizationConfig
 
 [[autodoc]] GGUFQuantizationConfig
+
+## QuantoConfig
+
+[[autodoc]] QuantoConfig
+
 ## TorchAoConfig
 
 [[autodoc]] TorchAoConfig

docs/source/en/api/schedulers/ddim_cogvideox.md

@@ -0,0 +1,19 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# CogVideoXDDIMScheduler
+
+`CogVideoXDDIMScheduler` is based on [Denoising Diffusion Implicit Models](https://huggingface.co/papers/2010.02502), specifically for CogVideoX models.
+
+## CogVideoXDDIMScheduler
+
+[[autodoc]] CogVideoXDDIMScheduler

docs/source/en/api/schedulers/multistep_dpm_solver_cogvideox.md

@@ -0,0 +1,19 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# CogVideoXDPMScheduler
+
+`CogVideoXDPMScheduler` is based on [DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps](https://huggingface.co/papers/2206.00927) and [DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models](https://huggingface.co/papers/2211.01095), specifically for CogVideoX models.
+
+## CogVideoXDPMScheduler
+
+[[autodoc]] CogVideoXDPMScheduler

docs/source/en/conceptual/evaluation.md

@@ -16,6 +16,11 @@ specific language governing permissions and limitations under the License.
     <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
 </a>
 
+> [!TIP]
+> This document has now grown outdated given the emergence of existing evaluation frameworks for diffusion models for image generation. Please check
+> out works like [HEIM](https://crfm.stanford.edu/helm/heim/latest/), [T2I-Compbench](https://arxiv.org/abs/2307.06350),
+> [GenEval](https://arxiv.org/abs/2310.11513).
+
 Evaluation of generative models like [Stable Diffusion](https://huggingface.co/docs/diffusers/stable_diffusion) is subjective in nature. But as practitioners and researchers, we often have to make careful choices amongst many different possibilities. So, when working with different generative models (like GANs, Diffusion, etc.), how do we choose one over the other?
 
 Qualitative evaluation of such models can be error-prone and might incorrectly influence a decision.

docs/source/en/hybrid_inference/api_reference.md

@@ -0,0 +1,5 @@
+# Hybrid Inference API Reference
+
+## Remote Decode
+
+[[autodoc]] utils.remote_utils.remote_decode

docs/source/en/hybrid_inference/overview.md

@@ -0,0 +1,54 @@
+<!--Copyright 2024 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Hybrid Inference
+
+**Empowering local AI builders with Hybrid Inference**
+
+
+> [!TIP]
+> Hybrid Inference is an [experimental feature](https://huggingface.co/blog/remote_vae).
+> Feedback can be provided [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+
+## Why use Hybrid Inference?
+
+Hybrid Inference offers a fast and simple way to offload local generation requirements.
+
+- 🚀 **Reduced Requirements:** Access powerful models without expensive hardware.
+- 💎 **Without Compromise:** Achieve the highest quality without sacrificing performance.
+- 💰 **Cost Effective:** It's free! 🤑
+- 🎯 **Diverse Use Cases:** Fully compatible with Diffusers 🧨 and the wider community.
+- 🔧 **Developer-Friendly:** Simple requests, fast responses.
+
+---
+
+## Available Models
+
+* **VAE Decode 🖼️:** Quickly decode latent representations into high-quality images without compromising performance or workflow speed.
+* **VAE Encode 🔢 (coming soon):** Efficiently encode images into latent representations for generation and training.
+* **Text Encoders 📃 (coming soon):** Compute text embeddings for your prompts quickly and accurately, ensuring a smooth and high-quality workflow.
+
+---
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.
+
+## Contents
+
+The documentation is organized into two sections:
+
+* **VAE Decode** Learn the basics of how to use VAE Decode with Hybrid Inference.
+* **API Reference** Dive into task-specific settings and parameters.

docs/source/en/hybrid_inference/vae_decode.md

@@ -0,0 +1,345 @@
+# Getting Started: VAE Decode with Hybrid Inference
+
+VAE decode is an essential component of diffusion models - turning latent representations into images or videos.
+
+## Memory
+
+These tables demonstrate the VRAM requirements for VAE decode with SD v1 and SD XL on different GPUs.
+
+For the majority of these GPUs the memory usage % dictates other models (text encoders, UNet/Transformer) must be offloaded, or tiled decoding has to be used which increases time taken and impacts quality.
+
+<details><summary>SD v1.5</summary>
+
+| GPU | Resolution | Time (seconds) | Memory (%) | Tiled Time (secs) | Tiled Memory (%) |
+| --- | --- | --- | --- | --- | --- |
+| NVIDIA GeForce RTX 4090 | 512x512 | 0.031 | 5.60% | 0.031 (0%) | 5.60% |
+| NVIDIA GeForce RTX 4090 | 1024x1024 | 0.148 | 20.00% | 0.301 (+103%) | 5.60% |
+| NVIDIA GeForce RTX 4080 | 512x512 | 0.05 | 8.40% | 0.050 (0%) | 8.40% |
+| NVIDIA GeForce RTX 4080 | 1024x1024 | 0.224 | 30.00% | 0.356 (+59%) | 8.40% |
+| NVIDIA GeForce RTX 4070 Ti | 512x512 | 0.066 | 11.30% | 0.066 (0%) | 11.30% |
+| NVIDIA GeForce RTX 4070 Ti | 1024x1024 | 0.284 | 40.50% | 0.454 (+60%) | 11.40% |
+| NVIDIA GeForce RTX 3090 | 512x512 | 0.062 | 5.20% | 0.062 (0%) | 5.20% |
+| NVIDIA GeForce RTX 3090 | 1024x1024 | 0.253 | 18.50% | 0.464 (+83%) | 5.20% |
+| NVIDIA GeForce RTX 3080 | 512x512 | 0.07 | 12.80% | 0.070 (0%) | 12.80% |
+| NVIDIA GeForce RTX 3080 | 1024x1024 | 0.286 | 45.30% | 0.466 (+63%) | 12.90% |
+| NVIDIA GeForce RTX 3070 | 512x512 | 0.102 | 15.90% | 0.102 (0%) | 15.90% |
+| NVIDIA GeForce RTX 3070 | 1024x1024 | 0.421 | 56.30% | 0.746 (+77%) | 16.00% |
+
+</details>
+
+<details><summary>SDXL</summary>
+
+| GPU | Resolution | Time (seconds) | Memory Consumed (%) | Tiled Time (seconds) | Tiled Memory (%) |
+| --- | --- | --- | --- | --- | --- |
+| NVIDIA GeForce RTX 4090 | 512x512 | 0.057 | 10.00% | 0.057 (0%) | 10.00% |
+| NVIDIA GeForce RTX 4090 | 1024x1024 | 0.256 | 35.50% | 0.257 (+0.4%) | 35.50% |
+| NVIDIA GeForce RTX 4080 | 512x512 | 0.092 | 15.00% | 0.092 (0%) | 15.00% |
+| NVIDIA GeForce RTX 4080 | 1024x1024 | 0.406 | 53.30% | 0.406 (0%) | 53.30% |
+| NVIDIA GeForce RTX 4070 Ti | 512x512 | 0.121 | 20.20% | 0.120 (-0.8%) | 20.20% |
+| NVIDIA GeForce RTX 4070 Ti | 1024x1024 | 0.519 | 72.00% | 0.519 (0%) | 72.00% |
+| NVIDIA GeForce RTX 3090 | 512x512 | 0.107 | 10.50% | 0.107 (0%) | 10.50% |
+| NVIDIA GeForce RTX 3090 | 1024x1024 | 0.459 | 38.00% | 0.460 (+0.2%) | 38.00% |
+| NVIDIA GeForce RTX 3080 | 512x512 | 0.121 | 25.60% | 0.121 (0%) | 25.60% |
+| NVIDIA GeForce RTX 3080 | 1024x1024 | 0.524 | 93.00% | 0.524 (0%) | 93.00% |
+| NVIDIA GeForce RTX 3070 | 512x512 | 0.183 | 31.80% | 0.183 (0%) | 31.80% |
+| NVIDIA GeForce RTX 3070 | 1024x1024 | 0.794 | 96.40% | 0.794 (0%) | 96.40% |
+
+</details>
+
+## Available VAEs
+
+|   | **Endpoint** | **Model** |
+|:-:|:-----------:|:--------:|
+| **Stable Diffusion v1** | [https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud](https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud) | [`stabilityai/sd-vae-ft-mse`](https://hf.co/stabilityai/sd-vae-ft-mse) |
+| **Stable Diffusion XL** | [https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud](https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud) | [`madebyollin/sdxl-vae-fp16-fix`](https://hf.co/madebyollin/sdxl-vae-fp16-fix) |
+| **Flux** | [https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud](https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud) | [`black-forest-labs/FLUX.1-schnell`](https://hf.co/black-forest-labs/FLUX.1-schnell) |
+| **HunyuanVideo** | [https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud](https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud) | [`hunyuanvideo-community/HunyuanVideo`](https://hf.co/hunyuanvideo-community/HunyuanVideo) |
+
+
+> [!TIP]
+> Model support can be requested [here](https://github.com/huggingface/diffusers/issues/new?template=remote-vae-pilot-feedback.yml).
+
+
+## Code
+
+> [!TIP]
+> Install `diffusers` from `main` to run the code: `pip install git+https://github.com/huggingface/diffusers@main`
+
+
+A helper method simplifies interacting with Hybrid Inference.
+
+```python
+from diffusers.utils.remote_utils import remote_decode
+```
+
+### Basic example
+
+Here, we show how to use the remote VAE on random tensors.
+
+<details><summary>Code</summary>
+
+```python
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=torch.randn([1, 4, 64, 64], dtype=torch.float16),
+    scaling_factor=0.18215,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/output.png"/>
+</figure>
+
+Usage for Flux is slightly different. Flux latents are packed so we need to send the `height` and `width`.
+
+<details><summary>Code</summary>
+
+```python
+image = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=torch.randn([1, 4096, 64], dtype=torch.float16),
+    height=1024,
+    width=1024,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/flux_random_latent.png"/>
+</figure>
+
+Finally, an example for HunyuanVideo.
+
+<details><summary>Code</summary>
+
+```python
+video = remote_decode(
+    endpoint="https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=torch.randn([1, 16, 3, 40, 64], dtype=torch.float16),
+    output_type="mp4",
+)
+with open("video.mp4", "wb") as f:
+    f.write(video)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+   <video
+      alt="queue.mp4"
+      autoplay loop autobuffer muted playsinline
+    >
+    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/video_1.mp4" type="video/mp4">
+  </video>
+</figure>
+
+
+### Generation
+
+But we want to use the VAE on an actual pipeline to get an actual image, not random noise. The example below shows how to do it with SD v1.5. 
+
+<details><summary>Code</summary>
+
+```python
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    variant="fp16",
+    vae=None,
+).to("cuda")
+
+prompt = "Strawberry ice cream, in a stylish modern glass, coconut, splashing milk cream and honey, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious"
+
+latent = pipe(
+    prompt=prompt,
+    output_type="latent",
+).images
+image = remote_decode(
+    endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    scaling_factor=0.18215,
+)
+image.save("test.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/test.jpg"/>
+</figure>
+
+Here’s another example with Flux.
+
+<details><summary>Code</summary>
+
+```python
+from diffusers import FluxPipeline
+
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-schnell",
+    torch_dtype=torch.bfloat16,
+    vae=None,
+).to("cuda")
+
+prompt = "Strawberry ice cream, in a stylish modern glass, coconut, splashing milk cream and honey, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious"
+
+latent = pipe(
+    prompt=prompt,
+    guidance_scale=0.0,
+    num_inference_steps=4,
+    output_type="latent",
+).images
+image = remote_decode(
+    endpoint="https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    height=1024,
+    width=1024,
+    scaling_factor=0.3611,
+    shift_factor=0.1159,
+)
+image.save("test.jpg")
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/test_1.jpg"/>
+</figure>
+
+Here’s an example with HunyuanVideo.
+
+<details><summary>Code</summary>
+
+```python
+from diffusers import HunyuanVideoPipeline, HunyuanVideoTransformer3DModel
+
+model_id = "hunyuanvideo-community/HunyuanVideo"
+transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+    model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+)
+pipe = HunyuanVideoPipeline.from_pretrained(
+    model_id, transformer=transformer, vae=None, torch_dtype=torch.float16
+).to("cuda")
+
+latent = pipe(
+    prompt="A cat walks on the grass, realistic",
+    height=320,
+    width=512,
+    num_frames=61,
+    num_inference_steps=30,
+    output_type="latent",
+).frames
+
+video = remote_decode(
+    endpoint="https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/",
+    tensor=latent,
+    output_type="mp4",
+)
+
+if isinstance(video, bytes):
+    with open("video.mp4", "wb") as f:
+        f.write(video)
+```
+
+</details>
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+   <video
+      alt="queue.mp4"
+      autoplay loop autobuffer muted playsinline
+    >
+    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/video.mp4" type="video/mp4">
+  </video>
+</figure>
+
+
+### Queueing
+
+One of the great benefits of using a remote VAE is that we can queue multiple generation requests. While the current latent is being processed for decoding, we can already queue another one. This helps improve concurrency. 
+
+
+<details><summary>Code</summary>
+
+```python
+import queue
+import threading
+from IPython.display import display
+from diffusers import StableDiffusionPipeline
+
+def decode_worker(q: queue.Queue):
+    while True:
+        item = q.get()
+        if item is None:
+            break
+        image = remote_decode(
+            endpoint="https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/",
+            tensor=item,
+            scaling_factor=0.18215,
+        )
+        display(image)
+        q.task_done()
+
+q = queue.Queue()
+thread = threading.Thread(target=decode_worker, args=(q,), daemon=True)
+thread.start()
+
+def decode(latent: torch.Tensor):
+    q.put(latent)
+
+prompts = [
+    "Blueberry ice cream, in a stylish modern glass , ice cubes, nuts, mint leaves, splashing milk cream, in a gradient purple background, fluid motion, dynamic movement, cinematic lighting, Mysterious",
+    "Lemonade in a glass, mint leaves, in an aqua and white background, flowers, ice cubes, halo, fluid motion, dynamic movement, soft lighting, digital painting, rule of thirds composition, Art by Greg rutkowski, Coby whitmore",
+    "Comic book art, beautiful, vintage, pastel neon colors, extremely detailed pupils, delicate features, light on face, slight smile, Artgerm, Mary Blair, Edmund Dulac, long dark locks, bangs, glowing, fashionable style, fairytale ambience, hot pink.",
+    "Masterpiece, vanilla cone ice cream garnished with chocolate syrup, crushed nuts, choco flakes, in a brown background, gold, cinematic lighting, Art by WLOP",
+    "A bowl of milk, falling cornflakes, berries, blueberries, in a white background, soft lighting, intricate details, rule of thirds, octane render, volumetric lighting",
+    "Cold Coffee with cream, crushed almonds, in a glass, choco flakes, ice cubes, wet, in a wooden background, cinematic lighting, hyper realistic painting, art by Carne Griffiths, octane render, volumetric lighting, fluid motion, dynamic movement, muted colors,",
+]
+
+pipe = StableDiffusionPipeline.from_pretrained(
+    "Lykon/dreamshaper-8",
+    torch_dtype=torch.float16,
+    vae=None,
+).to("cuda")
+
+pipe.unet = pipe.unet.to(memory_format=torch.channels_last)
+pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+_ = pipe(
+    prompt=prompts[0],
+    output_type="latent",
+)
+
+for prompt in prompts:
+    latent = pipe(
+        prompt=prompt,
+        output_type="latent",
+    ).images
+    decode(latent)
+
+q.put(None)
+thread.join()
+```
+
+</details>
+
+
+<figure class="image flex flex-col items-center justify-center text-center m-0 w-full">
+   <video
+      alt="queue.mp4"
+      autoplay loop autobuffer muted playsinline
+    >
+    <source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/remote_vae/queue.mp4" type="video/mp4">
+  </video>
+</figure>
+
+## Integrations
+
+* **[SD.Next](https://github.com/vladmandic/sdnext):** All-in-one UI with direct supports Hybrid Inference.
+* **[ComfyUI-HFRemoteVae](https://github.com/kijai/ComfyUI-HFRemoteVae):** ComfyUI node for Hybrid Inference.

docs/source/en/quantization/overview.md

@@ -36,5 +36,6 @@ Diffusers currently supports the following quantization methods.
 - [BitsandBytes](./bitsandbytes)
 - [TorchAO](./torchao)
 - [GGUF](./gguf)
+- [Quanto](./quanto.md)
 
 [This resource](https://huggingface.co/docs/transformers/main/en/quantization/overview#when-to-use-what) provides a good overview of the pros and cons of different quantization techniques.

docs/source/en/quantization/quanto.md

@@ -0,0 +1,148 @@
+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+-->
+
+# Quanto
+
+[Quanto](https://github.com/huggingface/optimum-quanto) is a PyTorch quantization backend for [Optimum](https://huggingface.co/docs/optimum/en/index). It has been designed with versatility and simplicity in mind:
+
+- All features are available in eager mode (works with non-traceable models)
+- Supports quantization aware training
+- Quantized models are compatible with `torch.compile`
+- Quantized models are Device agnostic (e.g CUDA,XPU,MPS,CPU)
+
+In order to use the Quanto backend, you will first need to install `optimum-quanto>=0.2.6` and `accelerate`
+
+```shell
+pip install optimum-quanto accelerate
+```
+
+Now you can quantize a model by passing the `QuantoConfig` object to the `from_pretrained()` method. Although the Quanto library does allow quantizing `nn.Conv2d` and `nn.LayerNorm` modules, currently, Diffusers only supports quantizing the weights in the `nn.Linear` layers of a model. The following snippet demonstrates how to apply `float8` quantization with Quanto.   
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="float8")
+transformer = FluxTransformer2DModel.from_pretrained(
+      model_id,
+      subfolder="transformer",
+      quantization_config=quantization_config,
+      torch_dtype=torch.bfloat16,
+)
+
+pipe = FluxPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch_dtype)
+pipe.to("cuda")
+
+prompt = "A cat holding a sign that says hello world"
+image = pipe(
+    prompt, num_inference_steps=50, guidance_scale=4.5, max_sequence_length=512
+).images[0]
+image.save("output.png")
+```
+
+## Skipping Quantization on specific modules
+
+It is possible to skip applying quantization on certain modules using the `modules_to_not_convert` argument in the `QuantoConfig`. Please ensure that the modules passed in to this argument match the keys of the modules in the `state_dict`  
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="float8", modules_to_not_convert=["proj_out"])
+transformer = FluxTransformer2DModel.from_pretrained(
+      model_id,
+      subfolder="transformer",
+      quantization_config=quantization_config,
+      torch_dtype=torch.bfloat16,
+)
+```
+
+## Using `from_single_file` with the Quanto Backend
+
+`QuantoConfig` is compatible with `~FromOriginalModelMixin.from_single_file`. 
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+ckpt_path = "https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors"
+quantization_config = QuantoConfig(weights_dtype="float8")
+transformer = FluxTransformer2DModel.from_single_file(ckpt_path, quantization_config=quantization_config, torch_dtype=torch.bfloat16)
+```
+
+## Saving Quantized models
+
+Diffusers supports serializing Quanto models using the `~ModelMixin.save_pretrained` method.
+
+The serialization and loading requirements are different for models quantized directly with the Quanto library and models quantized
+with Diffusers using Quanto as the backend. It is currently not possible to load models quantized directly with Quanto into Diffusers using `~ModelMixin.from_pretrained`
+
+```python
+import torch
+from diffusers import FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="float8")
+transformer = FluxTransformer2DModel.from_pretrained(
+      model_id,
+      subfolder="transformer",
+      quantization_config=quantization_config,
+      torch_dtype=torch.bfloat16,
+)
+# save quantized model to reuse
+transformer.save_pretrained("<your quantized model save path>")
+
+# you can reload your quantized model with
+model = FluxTransformer2DModel.from_pretrained("<your quantized model save path>")
+```
+
+## Using `torch.compile` with Quanto
+
+Currently the Quanto backend supports `torch.compile` for the following quantization types:
+
+- `int8` weights 
+
+```python
+import torch
+from diffusers import FluxPipeline, FluxTransformer2DModel, QuantoConfig
+
+model_id = "black-forest-labs/FLUX.1-dev"
+quantization_config = QuantoConfig(weights_dtype="int8")
+transformer = FluxTransformer2DModel.from_pretrained(
+    model_id,
+    subfolder="transformer",
+    quantization_config=quantization_config,
+    torch_dtype=torch.bfloat16,
+)
+transformer = torch.compile(transformer, mode="max-autotune", fullgraph=True)
+
+pipe = FluxPipeline.from_pretrained(
+    model_id, transformer=transformer, torch_dtype=torch_dtype
+)
+pipe.to("cuda")
+images = pipe("A cat holding a sign that says hello").images[0]
+images.save("flux-quanto-compile.png")
+```
+
+## Supported Quantization Types
+
+### Weights
+
+- float8
+- int8
+- int4
+- int2
+
+

docs/source/en/using-diffusers/callback.md

@@ -157,6 +157,84 @@ pipeline(
 )
 ```
 
+## IP Adapter Cutoff
+
+IP Adapter is an image prompt adapter that can be used for diffusion models without any changes to the underlying model. We can use the IP Adapter Cutoff Callback to disable the IP Adapter after a certain number of steps. To set up the callback, you need to specify the number of denoising steps after which the callback comes into effect. You can do so by using either one of these two arguments:
+
+- `cutoff_step_ratio`: Float number with the ratio of the steps.
+- `cutoff_step_index`: Integer number with the exact number of the step.
+
+We need to download the diffusion model and load the ip_adapter for it as follows:
+
+```py
+from diffusers import AutoPipelineForText2Image
+from diffusers.utils import load_image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter_sdxl.bin")
+pipeline.set_ip_adapter_scale(0.6)
+```
+The setup for the callback should look something like this:
+
+```py
+
+from diffusers import AutoPipelineForText2Image
+from diffusers.callbacks import IPAdapterScaleCutoffCallback
+from diffusers.utils import load_image
+import torch
+ 
+
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", 
+    torch_dtype=torch.float16
+).to("cuda")
+
+
+pipeline.load_ip_adapter(
+    "h94/IP-Adapter", 
+    subfolder="sdxl_models", 
+    weight_name="ip-adapter_sdxl.bin"
+)
+
+pipeline.set_ip_adapter_scale(0.6)
+
+
+callback = IPAdapterScaleCutoffCallback(
+    cutoff_step_ratio=None, 
+    cutoff_step_index=5
+)
+
+image = load_image(
+    "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ip_adapter_diner.png"
+)
+
+generator = torch.Generator(device="cuda").manual_seed(2628670641)
+
+images = pipeline(
+    prompt="a tiger sitting in a chair drinking orange juice",
+    ip_adapter_image=image,
+    negative_prompt="deformed, ugly, wrong proportion, low res, bad anatomy, worst quality, low quality",
+    generator=generator,
+    num_inference_steps=50,
+    callback_on_step_end=callback,
+).images
+
+images[0].save("custom_callback_img.png")
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/without_callback.png" alt="generated image of a tiger sitting in a chair drinking orange juice" />
+    <figcaption class="mt-2 text-center text-sm text-gray-500">without IPAdapterScaleCutoffCallback</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/with_callback2.png" alt="generated image of a tiger sitting in a chair drinking orange juice with ip adapter callback" />
+    <figcaption class="mt-2 text-center text-sm text-gray-500">with IPAdapterScaleCutoffCallback</figcaption>
+  </div>
+</div>
+
+
 ## Display image after each generation step
 
 > [!TIP]

docs/source/en/using-diffusers/marigold_usage.md

@@ -1,4 +1,6 @@
-<!--Copyright 2024 Marigold authors and The HuggingFace Team. All rights reserved.
+<!--
+Copyright 2023-2025 Marigold Team, ETH Zürich. All rights reserved.
+Copyright 2024-2025 The HuggingFace Team. All rights reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
 the License. You may obtain a copy of the License at
@@ -10,31 +12,38 @@ an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express o
 specific language governing permissions and limitations under the License.
 -->
 
-# Marigold Pipelines for Computer Vision Tasks
+# Marigold Computer Vision
 
-[Marigold](../api/pipelines/marigold) is a novel diffusion-based dense prediction approach, and a set of pipelines for various computer vision tasks, such as monocular depth estimation.
+**Marigold** is a diffusion-based [method](https://huggingface.co/papers/2312.02145) and a collection of [pipelines](../api/pipelines/marigold) designed for 
+dense computer vision tasks, including **monocular depth prediction**, **surface normals estimation**, and **intrinsic 
+image decomposition**.
 
-This guide will show you how to use Marigold to obtain fast and high-quality predictions for images and videos.
+This guide will walk you through using Marigold to generate fast and high-quality predictions for images and videos.
 
-Each pipeline supports one Computer Vision task, which takes an input RGB image as input and produces a *prediction* of the modality of interest, such as a depth map of the input image.
-Currently, the following tasks are implemented:
+Each pipeline is tailored for a specific computer vision task, processing an input RGB image and generating a 
+corresponding prediction.
+Currently, the following computer vision tasks are implemented:
 
-| Pipeline                                                                                                                                    | Predicted Modalities                                                                                             |                                                                       Demos                                                                        |
-|---------------------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------|:--------------------------------------------------------------------------------------------------------------------------------------------------:|
-| [MarigoldDepthPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_depth.py)     | [Depth](https://en.wikipedia.org/wiki/Depth_map), [Disparity](https://en.wikipedia.org/wiki/Binocular_disparity) | [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-lcm), [Slow Original Demo (DDIM)](https://huggingface.co/spaces/prs-eth/marigold) |
-| [MarigoldNormalsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_normals.py) | [Surface normals](https://en.wikipedia.org/wiki/Normal_mapping)                                                  |                                   [Fast Demo (LCM)](https://huggingface.co/spaces/prs-eth/marigold-normals-lcm)                                    |
+| Pipeline                                                                                                                                          | Recommended Model Checkpoints                                                                                                                                                                           |                              Spaces (Interactive Apps)                               | Predicted Modalities                                                                                                                                                               |
+|---------------------------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------:|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [MarigoldDepthPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_depth.py)           | [prs-eth/marigold-depth-v1-1](https://huggingface.co/prs-eth/marigold-depth-v1-1)                                                                                                                       |          [Depth Estimation](https://huggingface.co/spaces/prs-eth/marigold)          | [Depth](https://en.wikipedia.org/wiki/Depth_map), [Disparity](https://en.wikipedia.org/wiki/Binocular_disparity)                                                                   |
+| [MarigoldNormalsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_normals.py)       | [prs-eth/marigold-normals-v1-1](https://huggingface.co/prs-eth/marigold-normals-v1-1)                                                                                                                   | [Surface Normals Estimation](https://huggingface.co/spaces/prs-eth/marigold-normals) | [Surface normals](https://en.wikipedia.org/wiki/Normal_mapping)                                                                                                                    |
+| [MarigoldIntrinsicsPipeline](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/marigold/pipeline_marigold_intrinsics.py) | [prs-eth/marigold-iid-appearance-v1-1](https://huggingface.co/prs-eth/marigold-iid-appearance-v1-1),<br>[prs-eth/marigold-iid-lighting-v1-1](https://huggingface.co/prs-eth/marigold-iid-lighting-v1-1) | [Intrinsic Image Decomposition](https://huggingface.co/spaces/prs-eth/marigold-iid)  | [Albedo](https://en.wikipedia.org/wiki/Albedo), [Materials](https://www.n.aiq3d.com/wiki/roughnessmetalnessao-map), [Lighting](https://en.wikipedia.org/wiki/Diffuse_reflection)   |
 
-The original checkpoints can be found under the [PRS-ETH](https://huggingface.co/prs-eth/) Hugging Face organization.
-These checkpoints are meant to work with diffusers pipelines and the [original codebase](https://github.com/prs-eth/marigold).
-The original code can also be used to train new checkpoints.
+All original checkpoints are available under the [PRS-ETH](https://huggingface.co/prs-eth/) organization on Hugging Face.
+They are designed for use with diffusers pipelines and the [original codebase](https://github.com/prs-eth/marigold), which can also be used to train 
+new model checkpoints. 
+The following is a summary of the recommended checkpoints, all of which produce reliable results with 1 to 4 steps. 
 
-| Checkpoint                                                                                    | Modality | Comment                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
-|-----------------------------------------------------------------------------------------------|----------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| [prs-eth/marigold-v1-0](https://huggingface.co/prs-eth/marigold-v1-0)                         | Depth    | The first Marigold Depth checkpoint, which predicts *affine-invariant depth* maps. The performance of this checkpoint in benchmarks was studied in the original [paper](https://huggingface.co/papers/2312.02145). Designed to be used with the `DDIMScheduler` at inference, it requires at least 10 steps to get reliable predictions. Affine-invariant depth prediction has a range of values in each pixel between 0 (near plane) and 1 (far plane); both planes are chosen by the model as part of the inference process. See the `MarigoldImageProcessor` reference for visualization utilities. |
-| [prs-eth/marigold-depth-lcm-v1-0](https://huggingface.co/prs-eth/marigold-depth-lcm-v1-0)     | Depth    | The fast Marigold Depth checkpoint, fine-tuned from `prs-eth/marigold-v1-0`. Designed to be used with the `LCMScheduler` at inference, it requires as little as 1 step to get reliable predictions. The prediction reliability saturates at 4 steps and declines after that.                                                                                                                                                                                                                                                                                                                           |
-| [prs-eth/marigold-normals-v0-1](https://huggingface.co/prs-eth/marigold-normals-v0-1)         | Normals  | A preview checkpoint for the Marigold Normals pipeline. Designed to be used with the `DDIMScheduler` at inference, it requires at least 10 steps to get reliable predictions. The surface normals predictions are unit-length 3D vectors with values in the range from -1 to 1. *This checkpoint will be phased out after the release of `v1-0` version.*                                                                                                                                                                                                                                              |
-| [prs-eth/marigold-normals-lcm-v0-1](https://huggingface.co/prs-eth/marigold-normals-lcm-v0-1) | Normals  | The fast Marigold Normals checkpoint, fine-tuned from `prs-eth/marigold-normals-v0-1`. Designed to be used with the `LCMScheduler` at inference, it requires as little as 1 step to get reliable predictions. The prediction reliability saturates at 4 steps and declines after that. *This checkpoint will be phased out after the release of `v1-0` version.*                                                                                                                                                                                                                                       |
-The examples below are mostly given for depth prediction, but they can be universally applied with other supported modalities.
+| Checkpoint                                                                                          | Modality     | Comment                                                                                                                                                           |
+|-----------------------------------------------------------------------------------------------------|--------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| [prs-eth/marigold-depth-v1-1](https://huggingface.co/prs-eth/marigold-depth-v1-1)                   | Depth        | Affine-invariant depth prediction assigns each pixel a value between 0 (near plane) and 1 (far plane), with both planes determined by the model during inference. |
+| [prs-eth/marigold-normals-v0-1](https://huggingface.co/prs-eth/marigold-normals-v0-1)               | Normals      | The surface normals predictions are unit-length 3D vectors in the screen space camera, with values in the range from -1 to 1.                                     |
+| [prs-eth/marigold-iid-appearance-v1-1](https://huggingface.co/prs-eth/marigold-iid-appearance-v1-1) | Intrinsics   | InteriorVerse decomposition is comprised of Albedo and two BRDF material properties: Roughness and Metallicity.                                                   | 
+| [prs-eth/marigold-iid-lighting-v1-1](https://huggingface.co/prs-eth/marigold-iid-lighting-v1-1)     | Intrinsics   | HyperSim decomposition of an image \\(I\\) is comprised of Albedo \\(A\\), Diffuse shading \\(S\\), and Non-diffuse residual \\(R\\): \\(I = A*S+R\\).            | 
+
+The examples below are mostly given for depth prediction, but they can be universally applied to other supported 
+modalities.
 We showcase the predictions using the same input image of Albert Einstein generated by Midjourney.
 This makes it easier to compare visualizations of the predictions across various modalities and checkpoints.
 
@@ -47,19 +56,21 @@ This makes it easier to compare visualizations of the predictions across various
   </div>
 </div>
 
-### Depth Prediction Quick Start
+## Depth Prediction
 
-To get the first depth prediction, load `prs-eth/marigold-depth-lcm-v1-0` checkpoint into `MarigoldDepthPipeline` pipeline, put the image through the pipeline, and save the predictions:
+To get a depth prediction, load the `prs-eth/marigold-depth-v1-1` checkpoint into [`MarigoldDepthPipeline`], 
+put the image through the pipeline, and save the predictions:
 
 ```python
 import diffusers
 import torch
 
 pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
-    "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+    "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
 ).to("cuda")
 
 image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
 depth = pipe(image)
 
 vis = pipe.image_processor.visualize_depth(depth.prediction)
@@ -69,10 +80,13 @@ depth_16bit = pipe.image_processor.export_depth_to_16bit_png(depth.prediction)
 depth_16bit[0].save("einstein_depth_16bit.png")
 ```
 
-The visualization function for depth [`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_depth`] applies one of [matplotlib's colormaps](https://matplotlib.org/stable/users/explain/colors/colormaps.html) (`Spectral` by default) to map the predicted pixel values from a single-channel `[0, 1]` depth range into an RGB image.
-With the `Spectral` colormap, pixels with near depth are painted red, and far pixels are assigned blue color.
+The [`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_depth`] function applies one of 
+[matplotlib's colormaps](https://matplotlib.org/stable/users/explain/colors/colormaps.html) (`Spectral` by default) to map the predicted pixel values from a single-channel `[0, 1]` 
+depth range into an RGB image.
+With the `Spectral` colormap, pixels with near depth are painted red, and far pixels are blue.
 The 16-bit PNG file stores the single channel values mapped linearly from the `[0, 1]` range into `[0, 65535]`.
-Below are the raw and the visualized predictions; as can be seen, dark areas (mustache) are easier to distinguish in the visualization:
+Below are the raw and the visualized predictions. The darker and closer areas (mustache) are easier to distinguish in 
+the visualization.
 
 <div class="flex gap-4">
   <div style="flex: 1 1 50%; max-width: 50%;">
@@ -89,28 +103,33 @@ Below are the raw and the visualized predictions; as can be seen, dark areas (mu
   </div>
 </div>
 
-### Surface Normals Prediction Quick Start
+## Surface Normals Estimation
 
-Load `prs-eth/marigold-normals-lcm-v0-1` checkpoint into `MarigoldNormalsPipeline` pipeline, put the image through the pipeline, and save the predictions:
+Load the `prs-eth/marigold-normals-v1-1` checkpoint into [`MarigoldNormalsPipeline`], put the image through the 
+pipeline, and save the predictions:
 
 ```python
 import diffusers
 import torch
 
 pipe = diffusers.MarigoldNormalsPipeline.from_pretrained(
-    "prs-eth/marigold-normals-lcm-v0-1", variant="fp16", torch_dtype=torch.float16
+    "prs-eth/marigold-normals-v1-1", variant="fp16", torch_dtype=torch.float16
 ).to("cuda")
 
 image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
 normals = pipe(image)
 
 vis = pipe.image_processor.visualize_normals(normals.prediction)
 vis[0].save("einstein_normals.png")
 ```
 
-The visualization function for normals [`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_normals`] maps the three-dimensional prediction with pixel values in the range `[-1, 1]` into an RGB image.
-The visualization function supports flipping surface normals axes to make the visualization compatible with other choices of the frame of reference.
-Conceptually, each pixel is painted according to the surface normal vector in the frame of reference, where `X` axis points right, `Y` axis points up, and `Z` axis points at the viewer.
+The [`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_normals`] maps the three-dimensional 
+prediction with pixel values in the range `[-1, 1]` into an RGB image.
+The visualization function supports flipping surface normals axes to make the visualization compatible with other 
+choices of the frame of reference.
+Conceptually, each pixel is painted according to the surface normal vector in the frame of reference, where `X` axis 
+points right, `Y` axis points up, and `Z` axis points at the viewer.
 Below is the visualized prediction:
 
 <div class="flex gap-4" style="justify-content: center; width: 100%;">
@@ -122,25 +141,121 @@ Below is the visualized prediction:
   </div>
 </div>
 
-In this example, the nose tip almost certainly has a point on the surface, in which the surface normal vector points straight at the viewer, meaning that its coordinates are `[0, 0, 1]`.
+In this example, the nose tip almost certainly has a point on the surface, in which the surface normal vector points 
+straight at the viewer, meaning that its coordinates are `[0, 0, 1]`.
 This vector maps to the RGB `[128, 128, 255]`, which corresponds to the violet-blue color.
-Similarly, a surface normal on the cheek in the right part of the image has a large `X` component, which increases the red hue.
+Similarly, a surface normal on the cheek in the right part of the image has a large `X` component, which increases the 
+red hue.
 Points on the shoulders pointing up with a large `Y` promote green color.
 
-### Speeding up inference
+## Intrinsic Image Decomposition
 
-The above quick start snippets are already optimized for speed: they load the LCM checkpoint, use the `fp16` variant of weights and computation, and perform just one denoising diffusion step.
-The `pipe(image)` call completes in 280ms on RTX 3090 GPU.
-Internally, the input image is encoded with the Stable Diffusion VAE encoder, then the U-Net performs one denoising step, and finally, the prediction latent is decoded with the VAE decoder into pixel space.
-In this case, two out of three module calls are dedicated to converting between pixel and latent space of LDM.
-Because Marigold's latent space is compatible with the base Stable Diffusion, it is possible to speed up the pipeline call by more than 3x (85ms on RTX 3090) by using a [lightweight replacement of the SD VAE](../api/models/autoencoder_tiny):
+Marigold provides two models for Intrinsic Image Decomposition (IID): "Appearance" and "Lighting". 
+Each model produces Albedo maps, derived from InteriorVerse and Hypersim annotations, respectively.
+
+- The "Appearance" model also estimates Material properties: Roughness and Metallicity.
+- The "Lighting" model generates Diffuse Shading and Non-diffuse Residual.
+
+Here is the sample code saving predictions made by the "Appearance" model:
+
+```python
+import diffusers
+import torch
+
+pipe = diffusers.MarigoldIntrinsicsPipeline.from_pretrained(
+    "prs-eth/marigold-iid-appearance-v1-1", variant="fp16", torch_dtype=torch.float16
+).to("cuda")
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+intrinsics = pipe(image)
+
+vis = pipe.image_processor.visualize_intrinsics(intrinsics.prediction, pipe.target_properties)
+vis[0]["albedo"].save("einstein_albedo.png")
+vis[0]["roughness"].save("einstein_roughness.png")
+vis[0]["metallicity"].save("einstein_metallicity.png")
+```
+
+Another example demonstrating the predictions made by the "Lighting" model:
+
+```python
+import diffusers
+import torch
+
+pipe = diffusers.MarigoldIntrinsicsPipeline.from_pretrained(
+    "prs-eth/marigold-iid-lighting-v1-1", variant="fp16", torch_dtype=torch.float16
+).to("cuda")
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+intrinsics = pipe(image)
+
+vis = pipe.image_processor.visualize_intrinsics(intrinsics.prediction, pipe.target_properties)
+vis[0]["albedo"].save("einstein_albedo.png")
+vis[0]["shading"].save("einstein_shading.png")
+vis[0]["residual"].save("einstein_residual.png")
+```
+
+Both models share the same pipeline while supporting different decomposition types.
+The exact decomposition parameterization (e.g., sRGB vs. linear space) is stored in the 
+`pipe.target_properties` dictionary, which is passed into the 
+[`~pipelines.marigold.marigold_image_processing.MarigoldImageProcessor.visualize_intrinsics`] function.
+
+Below are some examples showcasing the predicted decomposition outputs. 
+All modalities can be inspected in the 
+[Intrinsic Image Decomposition](https://huggingface.co/spaces/prs-eth/marigold-iid) Space.
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/8c7986eaaab5eb9604eb88336311f46a7b0ff5ab/marigold/marigold_einstein_albedo.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Predicted albedo ("Appearance" model)
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 50%; max-width: 50%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/8c7986eaaab5eb9604eb88336311f46a7b0ff5ab/marigold/marigold_einstein_diffuse.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Predicted diffuse shading ("Lighting" model)
+    </figcaption>
+  </div>
+</div>
+
+## Speeding up inference
+
+The above quick start snippets are already optimized for quality and speed, loading the checkpoint, utilizing the 
+`fp16` variant of weights and computation, and performing the default number (4) of denoising diffusion steps.
+The first step to accelerate inference, at the expense of prediction quality, is to reduce the denoising diffusion 
+steps to the minimum:
 
 ```diff
   import diffusers
   import torch
 
   pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
-      "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+      "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
+  ).to("cuda")
+
+  image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+  
+- depth = pipe(image)
++ depth = pipe(image, num_inference_steps=1)
+```
+
+With this change, the `pipe` call completes in 280ms on RTX 3090 GPU.
+Internally, the input image is first encoded using the Stable Diffusion VAE encoder, followed by a single denoising 
+step performed by the U-Net. 
+Finally, the prediction latent is decoded with the VAE decoder into pixel space.
+In this setup, two out of three module calls are dedicated to converting between the pixel and latent spaces of the LDM.
+Since Marigold's latent space is compatible with Stable Diffusion 2.0, inference can be accelerated by more than 3x, 
+reducing the call time to 85ms on an RTX 3090, by using a [lightweight replacement of the SD VAE](../api/models/autoencoder_tiny). 
+Note that using a lightweight VAE may slightly reduce the visual quality of the predictions.
+
+```diff
+  import diffusers
+  import torch
+
+  pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+      "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
   ).to("cuda")
 
 + pipe.vae = diffusers.AutoencoderTiny.from_pretrained(
@@ -148,78 +263,77 @@ Because Marigold's latent space is compatible with the base Stable Diffusion, it
 + ).cuda()
 
   image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
-  depth = pipe(image)
+
+  depth = pipe(image, num_inference_steps=1)
 ```
 
-As suggested in [Optimizations](../optimization/torch2.0#torch.compile), adding `torch.compile` may squeeze extra performance depending on the target hardware:
+So far, we have optimized the number of diffusion steps and model components. Self-attention operations account for a 
+significant portion of computations. 
+Speeding them up can be achieved by using a more efficient attention processor:
 
 ```diff
   import diffusers
   import torch
++ from diffusers.models.attention_processor import AttnProcessor2_0
 
   pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
-      "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+      "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
   ).to("cuda")
 
++ pipe.vae.set_attn_processor(AttnProcessor2_0()) 
++ pipe.unet.set_attn_processor(AttnProcessor2_0())
+
+  image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+  depth = pipe(image, num_inference_steps=1)
+```
+
+Finally, as suggested in [Optimizations](../optimization/torch2.0#torch.compile), enabling `torch.compile` can further enhance performance depending on 
+the target hardware.
+However, compilation incurs a significant overhead during the first pipeline invocation, making it beneficial only when 
+the same pipeline instance is called repeatedly, such as within a loop.
+
+```diff
+  import diffusers
+  import torch
+  from diffusers.models.attention_processor import AttnProcessor2_0
+
+  pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+      "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
+  ).to("cuda")
+
+  pipe.vae.set_attn_processor(AttnProcessor2_0()) 
+  pipe.unet.set_attn_processor(AttnProcessor2_0())
+
++ pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=True)
 + pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
 
   image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
-  depth = pipe(image)
+
+  depth = pipe(image, num_inference_steps=1)
 ```
 
-## Qualitative Comparison with Depth Anything
-
-With the above speed optimizations, Marigold delivers predictions with more details and faster than [Depth Anything](https://huggingface.co/docs/transformers/main/en/model_doc/depth_anything) with the largest checkpoint [LiheYoung/depth-anything-large-hf](https://huggingface.co/LiheYoung/depth-anything-large-hf):
-
-<div class="flex gap-4">
-  <div style="flex: 1 1 50%; max-width: 50%;">
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_lcm_depth.png"/>
-    <figcaption class="mt-1 text-center text-sm text-gray-500">
-      Marigold LCM fp16 with Tiny AutoEncoder
-    </figcaption>
-  </div>
-  <div style="flex: 1 1 50%; max-width: 50%;">
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/einstein_depthanything_large.png"/>
-    <figcaption class="mt-1 text-center text-sm text-gray-500">
-      Depth Anything Large
-    </figcaption>
-  </div>
-</div>
-
 ## Maximizing Precision and Ensembling
 
 Marigold pipelines have a built-in ensembling mechanism combining multiple predictions from different random latents.
 This is a brute-force way of improving the precision of predictions, capitalizing on the generative nature of diffusion.
-The ensembling path is activated automatically when the `ensemble_size` argument is set greater than `1`.
+The ensembling path is activated automatically when the `ensemble_size` argument is set greater or equal than `3`.
 When aiming for maximum precision, it makes sense to adjust `num_inference_steps` simultaneously with `ensemble_size`.
 The recommended values vary across checkpoints but primarily depend on the scheduler type.
 The effect of ensembling is particularly well-seen with surface normals:
 
-```python
-import diffusers
+```diff
+  import diffusers
 
-model_path = "prs-eth/marigold-normals-v1-0"
+  pipe = diffusers.MarigoldNormalsPipeline.from_pretrained("prs-eth/marigold-normals-v1-1").to("cuda")
 
-model_paper_kwargs = {
-	diffusers.schedulers.DDIMScheduler: {
-		"num_inference_steps": 10,
-		"ensemble_size": 10,
-	},
-	diffusers.schedulers.LCMScheduler: {
-		"num_inference_steps": 4,
-		"ensemble_size": 5,
-	},
-}
+  image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
 
-image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+- depth = pipe(image)
++ depth = pipe(image, num_inference_steps=10, ensemble_size=5)
 
-pipe = diffusers.MarigoldNormalsPipeline.from_pretrained(model_path).to("cuda")
-pipe_kwargs = model_paper_kwargs[type(pipe.scheduler)]
-
-depth = pipe(image, **pipe_kwargs)
-
-vis = pipe.image_processor.visualize_normals(depth.prediction)
-vis[0].save("einstein_normals.png")
+  vis = pipe.image_processor.visualize_normals(depth.prediction)
+  vis[0].save("einstein_normals.png")
 ```
 
 <div class="flex gap-4">
@@ -237,93 +351,16 @@ vis[0].save("einstein_normals.png")
   </div>
 </div>
 
-As can be seen, all areas with fine-grained structurers, such as hair, got more conservative and on average more correct predictions.
+As can be seen, all areas with fine-grained structurers, such as hair, got more conservative and on average more 
+correct predictions.
 Such a result is more suitable for precision-sensitive downstream tasks, such as 3D reconstruction.
 
-## Quantitative Evaluation
-
-To evaluate Marigold quantitatively in standard leaderboards and benchmarks (such as NYU, KITTI, and other datasets), follow the evaluation protocol outlined in the paper: load the full precision fp32 model and use appropriate values for `num_inference_steps` and `ensemble_size`.
-Optionally seed randomness to ensure reproducibility. Maximizing `batch_size` will deliver maximum device utilization.
-
-```python
-import diffusers
-import torch
-
-device = "cuda"
-seed = 2024
-model_path = "prs-eth/marigold-v1-0"
-
-model_paper_kwargs = {
-	diffusers.schedulers.DDIMScheduler: {
-		"num_inference_steps": 50,
-		"ensemble_size": 10,
-	},
-	diffusers.schedulers.LCMScheduler: {
-		"num_inference_steps": 4,
-		"ensemble_size": 10,
-	},
-}
-
-image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
-
-generator = torch.Generator(device=device).manual_seed(seed)
-pipe = diffusers.MarigoldDepthPipeline.from_pretrained(model_path).to(device)
-pipe_kwargs = model_paper_kwargs[type(pipe.scheduler)]
-
-depth = pipe(image, generator=generator, **pipe_kwargs)
-
-# evaluate metrics
-```
-
-## Using Predictive Uncertainty
-
-The ensembling mechanism built into Marigold pipelines combines multiple predictions obtained from different random latents.
-As a side effect, it can be used to quantify epistemic (model) uncertainty; simply specify `ensemble_size` greater than 1 and set `output_uncertainty=True`.
-The resulting uncertainty will be available in the `uncertainty` field of the output.
-It can be visualized as follows:
-
-```python
-import diffusers
-import torch
-
-pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
-    "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
-).to("cuda")
-
-image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
-depth = pipe(
-	image,
-	ensemble_size=10,  # any number greater than 1; higher values yield higher precision
-	output_uncertainty=True,
-)
-
-uncertainty = pipe.image_processor.visualize_uncertainty(depth.uncertainty)
-uncertainty[0].save("einstein_depth_uncertainty.png")
-```
-
-<div class="flex gap-4">
-  <div style="flex: 1 1 50%; max-width: 50%;">
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_depth_uncertainty.png"/>
-    <figcaption class="mt-1 text-center text-sm text-gray-500">
-      Depth uncertainty
-    </figcaption>
-  </div>
-  <div style="flex: 1 1 50%; max-width: 50%;">
-    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_normals_uncertainty.png"/>
-    <figcaption class="mt-1 text-center text-sm text-gray-500">
-      Surface normals uncertainty
-    </figcaption>
-  </div>
-</div>
-
-The interpretation of uncertainty is easy: higher values (white) correspond to pixels, where the model struggles to make consistent predictions.
-Evidently, the depth model is the least confident around edges with discontinuity, where the object depth changes drastically.
-The surface normals model is the least confident in fine-grained structures, such as hair, and dark areas, such as the collar.
-
 ## Frame-by-frame Video Processing with Temporal Consistency
 
-Due to Marigold's generative nature, each prediction is unique and defined by the random noise sampled for the latent initialization.
-This becomes an obvious drawback compared to traditional end-to-end dense regression networks, as exemplified in the following videos:
+Due to Marigold's generative nature, each prediction is unique and defined by the random noise sampled for the latent 
+initialization.
+This becomes an obvious drawback compared to traditional end-to-end dense regression networks, as exemplified in the 
+following videos:
 
 <div class="flex gap-4">
   <div style="flex: 1 1 50%; max-width: 50%;">
@@ -336,26 +373,32 @@ This becomes an obvious drawback compared to traditional end-to-end dense regres
   </div>
 </div>
 
-To address this issue, it is possible to pass `latents` argument to the pipelines, which defines the starting point of diffusion.
-Empirically, we found that a convex combination of the very same starting point noise latent and the latent corresponding to the previous frame prediction give sufficiently smooth results, as implemented in the snippet below:
+To address this issue, it is possible to pass `latents` argument to the pipelines, which defines the starting point of 
+diffusion.
+Empirically, we found that a convex combination of the very same starting point noise latent and the latent 
+corresponding to the previous frame prediction give sufficiently smooth results, as implemented in the snippet below:
 
 ```python
 import imageio
-from PIL import Image
-from tqdm import tqdm
 import diffusers
 import torch
+from diffusers.models.attention_processor import AttnProcessor2_0
+from PIL import Image
+from tqdm import tqdm
 
 device = "cuda"
-path_in = "obama.mp4"
+path_in = "https://huggingface.co/spaces/prs-eth/marigold-lcm/resolve/c7adb5427947d2680944f898cd91d386bf0d4924/files/video/obama.mp4"
 path_out = "obama_depth.gif"
 
 pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
-    "prs-eth/marigold-depth-lcm-v1-0", variant="fp16", torch_dtype=torch.float16
+    "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
 ).to(device)
 pipe.vae = diffusers.AutoencoderTiny.from_pretrained(
     "madebyollin/taesd", torch_dtype=torch.float16
 ).to(device)
+pipe.unet.set_attn_processor(AttnProcessor2_0())
+pipe.vae = torch.compile(pipe.vae, mode="reduce-overhead", fullgraph=True)
+pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
 pipe.set_progress_bar_config(disable=True)
 
 with imageio.get_reader(path_in) as reader:
@@ -373,7 +416,11 @@ with imageio.get_reader(path_in) as reader:
             latents = 0.9 * latents + 0.1 * last_frame_latent
 
         depth = pipe(
-			frame, match_input_resolution=False, latents=latents, output_latent=True
+            frame,
+            num_inference_steps=1,
+            match_input_resolution=False, 
+            latents=latents, 
+            output_latent=True,
         )
         last_frame_latent = depth.latent
         out.append(pipe.image_processor.visualize_depth(depth.prediction)[0])
@@ -382,7 +429,8 @@ with imageio.get_reader(path_in) as reader:
 ```
 
 Here, the diffusion process starts from the given computed latent.
-The pipeline sets `output_latent=True` to access `out.latent` and computes its contribution to the next frame's latent initialization.
+The pipeline sets `output_latent=True` to access `out.latent` and computes its contribution to the next frame's latent 
+initialization.
 The result is much more stable now:
 
 <div class="flex gap-4">
@@ -414,7 +462,7 @@ image = diffusers.utils.load_image(
 )
 
 pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
-    "prs-eth/marigold-depth-lcm-v1-0", torch_dtype=torch.float16, variant="fp16"
+    "prs-eth/marigold-depth-v1-1", torch_dtype=torch.float16, variant="fp16"
 ).to(device)
 
 depth_image = pipe(image, generator=generator).prediction
@@ -463,4 +511,95 @@ controlnet_out[0].save("motorcycle_controlnet_out.png")
   </div>
 </div>
 
-Hopefully, you will find Marigold useful for solving your downstream tasks, be it a part of a more broad generative workflow, or a perception task, such as 3D reconstruction.
+## Quantitative Evaluation
+
+To evaluate Marigold quantitatively in standard leaderboards and benchmarks (such as NYU, KITTI, and other datasets), 
+follow the evaluation protocol outlined in the paper: load the full precision fp32 model and use appropriate values 
+for `num_inference_steps` and `ensemble_size`.
+Optionally seed randomness to ensure reproducibility. 
+Maximizing `batch_size` will deliver maximum device utilization.
+
+```python
+import diffusers
+import torch
+
+device = "cuda"
+seed = 2024
+
+generator = torch.Generator(device=device).manual_seed(seed)
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained("prs-eth/marigold-depth-v1-1").to(device)
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+depth = pipe(
+    image, 
+    num_inference_steps=4,  # set according to the evaluation protocol from the paper
+    ensemble_size=10,       # set according to the evaluation protocol from the paper
+    generator=generator,
+)
+
+# evaluate metrics
+```
+
+## Using Predictive Uncertainty
+
+The ensembling mechanism built into Marigold pipelines combines multiple predictions obtained from different random 
+latents.
+As a side effect, it can be used to quantify epistemic (model) uncertainty; simply specify `ensemble_size` greater 
+or equal than 3 and set `output_uncertainty=True`.
+The resulting uncertainty will be available in the `uncertainty` field of the output.
+It can be visualized as follows:
+
+```python
+import diffusers
+import torch
+
+pipe = diffusers.MarigoldDepthPipeline.from_pretrained(
+    "prs-eth/marigold-depth-v1-1", variant="fp16", torch_dtype=torch.float16
+).to("cuda")
+
+image = diffusers.utils.load_image("https://marigoldmonodepth.github.io/images/einstein.jpg")
+
+depth = pipe(
+	image,
+	ensemble_size=10,  # any number >= 3
+	output_uncertainty=True,
+)
+
+uncertainty = pipe.image_processor.visualize_uncertainty(depth.uncertainty)
+uncertainty[0].save("einstein_depth_uncertainty.png")
+```
+
+<div class="flex gap-4">
+  <div style="flex: 1 1 33%; max-width: 33%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_depth_uncertainty.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Depth uncertainty
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 33%; max-width: 33%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/marigold/marigold_einstein_normals_uncertainty.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Surface normals uncertainty
+    </figcaption>
+  </div>
+  <div style="flex: 1 1 33%; max-width: 33%;">
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/4f83035d84a24e5ec44fdda129b1d51eba12ce04/marigold/marigold_einstein_albedo_uncertainty.png"/>
+    <figcaption class="mt-1 text-center text-sm text-gray-500">
+      Albedo uncertainty
+    </figcaption>
+  </div>
+</div>
+
+The interpretation of uncertainty is easy: higher values (white) correspond to pixels, where the model struggles to 
+make consistent predictions.
+- The depth model exhibits the most uncertainty around discontinuities, where object depth changes abruptly.
+- The surface normals model is least confident in fine-grained structures like hair and in dark regions such as the 
+collar area.
+- Albedo uncertainty is represented as an RGB image, as it captures uncertainty independently for each color channel, 
+unlike depth and surface normals. It is also higher in shaded regions and at discontinuities.
+
+## Conclusion
+
+We hope Marigold proves valuable for your downstream tasks, whether as part of a broader generative workflow or for 
+perception-based applications like 3D reconstruction.

docs/source/en/using-diffusers/weighted_prompts.md

@@ -215,7 +215,7 @@ image
 
 Prompt weighting provides a way to emphasize or de-emphasize certain parts of a prompt, allowing for more control over the generated image. A prompt can include several concepts, which gets turned into contextualized text embeddings. The embeddings are used by the model to condition its cross-attention layers to generate an image (read the Stable Diffusion [blog post](https://huggingface.co/blog/stable_diffusion) to learn more about how it works).
 
-Prompt weighting works by increasing or decreasing the scale of the text embedding vector that corresponds to its concept in the prompt because you may not necessarily want the model to focus on all concepts equally. The easiest way to prepare the prompt-weighted embeddings is to use [Compel](https://github.com/damian0815/compel), a text prompt-weighting and blending library. Once you have the prompt-weighted embeddings, you can pass them to any pipeline that has a [`prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.__call__.prompt_embeds) (and optionally [`negative_prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.__call__.negative_prompt_embeds)) parameter, such as [`StableDiffusionPipeline`], [`StableDiffusionControlNetPipeline`], and [`StableDiffusionXLPipeline`].
+Prompt weighting works by increasing or decreasing the scale of the text embedding vector that corresponds to its concept in the prompt because you may not necessarily want the model to focus on all concepts equally. The easiest way to prepare the prompt embeddings is to use [Stable Diffusion Long Prompt Weighted Embedding](https://github.com/xhinker/sd_embed) (sd_embed). Once you have the prompt-weighted embeddings, you can pass them to any pipeline that has a [prompt_embeds](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.__call__.prompt_embeds) (and optionally [negative_prompt_embeds](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/text2img#diffusers.StableDiffusionPipeline.__call__.negative_prompt_embeds)) parameter, such as [`StableDiffusionPipeline`], [`StableDiffusionControlNetPipeline`], and [`StableDiffusionXLPipeline`].
 
 <Tip>
 
@@ -223,136 +223,99 @@ If your favorite pipeline doesn't have a `prompt_embeds` parameter, please open
 
 </Tip>
 
-This guide will show you how to weight and blend your prompts with Compel in 🤗 Diffusers.
+This guide will show you how to weight your prompts with sd_embed.
 
-Before you begin, make sure you have the latest version of Compel installed:
+Before you begin, make sure you have the latest version of sd_embed installed:
 
-```py
-# uncomment to install in Colab
-#!pip install compel --upgrade
+```bash
+pip install git+https://github.com/xhinker/sd_embed.git@main
 ```
 
-For this guide, let's generate an image with the prompt `"a red cat playing with a ball"` using the [`StableDiffusionPipeline`]:
+For this example, let's use [`StableDiffusionXLPipeline`].
 
 ```py
-from diffusers import StableDiffusionPipeline, UniPCMultistepScheduler
+from diffusers import StableDiffusionXLPipeline, UniPCMultistepScheduler
 import torch
 
-pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", use_safetensors=True)
+pipe = StableDiffusionXLPipeline.from_pretrained("Lykon/dreamshaper-xl-1-0", torch_dtype=torch.float16)
 pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
 pipe.to("cuda")
+```
 
-prompt = "a red cat playing with a ball"
+To upweight or downweight a concept, surround the text with parentheses. More parentheses applies a heavier weight on the text. You can also append a numerical multiplier to the text to indicate how much you want to increase or decrease its weights by.
 
-generator = torch.Generator(device="cpu").manual_seed(33)
+| format | multiplier |
+|---|---|
+| `(hippo)` | increase by 1.1x |
+| `((hippo))` | increase by 1.21x |
+| `(hippo:1.5)` | increase by 1.5x |
+| `(hippo:0.5)` | decrease by 4x |
 
-image = pipe(prompt, generator=generator, num_inference_steps=20).images[0]
+Create a prompt and use a combination of parentheses and numerical multipliers to upweight various text.
+
+```py
+from sd_embed.embedding_funcs import get_weighted_text_embeddings_sdxl
+
+prompt = """A whimsical and creative image depicting a hybrid creature that is a mix of a waffle and a hippopotamus. 
+This imaginative creature features the distinctive, bulky body of a hippo, 
+but with a texture and appearance resembling a golden-brown, crispy waffle. 
+The creature might have elements like waffle squares across its skin and a syrup-like sheen. 
+It's set in a surreal environment that playfully combines a natural water habitat of a hippo with elements of a breakfast table setting, 
+possibly including oversized utensils or plates in the background. 
+The image should evoke a sense of playful absurdity and culinary fantasy.
+"""
+
+neg_prompt = """\
+skin spots,acnes,skin blemishes,age spot,(ugly:1.2),(duplicate:1.2),(morbid:1.21),(mutilated:1.2),\
+(tranny:1.2),mutated hands,(poorly drawn hands:1.5),blurry,(bad anatomy:1.2),(bad proportions:1.3),\
+extra limbs,(disfigured:1.2),(missing arms:1.2),(extra legs:1.2),(fused fingers:1.5),\
+(too many fingers:1.5),(unclear eyes:1.2),lowers,bad hands,missing fingers,extra digit,\
+bad hands,missing fingers,(extra arms and legs),(worst quality:2),(low quality:2),\
+(normal quality:2),lowres,((monochrome)),((grayscale))
+"""
+```
+
+Use the `get_weighted_text_embeddings_sdxl` function to generate the prompt embeddings and the negative prompt embeddings. It'll also generated the pooled and negative pooled prompt embeddings since you're using the SDXL model.
+
+> [!TIP]
+> You can safely ignore the error message below about the token index length exceeding the models maximum sequence length. All your tokens will be used in the embedding process.
+>
+> ```
+> Token indices sequence length is longer than the specified maximum sequence length for this model
+> ```
+
+```py
+( 
+  prompt_embeds,
+  prompt_neg_embeds,
+  pooled_prompt_embeds,
+  negative_pooled_prompt_embeds
+) = get_weighted_text_embeddings_sdxl(
+    pipe,
+    prompt=prompt,
+    neg_prompt=neg_prompt
+)
+
+image = pipe(
+    prompt_embeds=prompt_embeds,
+    negative_prompt_embeds=prompt_neg_embeds,
+    pooled_prompt_embeds=pooled_prompt_embeds,
+    negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+    num_inference_steps=30,
+    height=1024,
+    width=1024 + 512,
+    guidance_scale=4.0,
+    generator=torch.Generator("cuda").manual_seed(2)
+).images[0]
 image
 ```
 
 <div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/compel/forest_0.png"/>
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sd_embed_sdxl.png"/>
 </div>
 
-### Weighting
-
-You'll notice there is no "ball" in the image! Let's use compel to upweight the concept of "ball" in the prompt. Create a [`Compel`](https://github.com/damian0815/compel/blob/main/doc/compel.md#compel-objects) object, and pass it a tokenizer and text encoder:
-
-```py
-from compel import Compel
-
-compel_proc = Compel(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder)
-```
-
-compel uses `+` or `-` to increase or decrease the weight of a word in the prompt. To increase the weight of "ball":
-
-<Tip>
-
-`+` corresponds to the value `1.1`, `++` corresponds to `1.1^2`, and so on. Similarly, `-` corresponds to `0.9` and `--` corresponds to `0.9^2`. Feel free to experiment with adding more `+` or `-` in your prompt!
-
-</Tip>
-
-```py
-prompt = "a red cat playing with a ball++"
-```
-
-Pass the prompt to `compel_proc` to create the new prompt embeddings which are passed to the pipeline:
-
-```py
-prompt_embeds = compel_proc(prompt)
-generator = torch.manual_seed(33)
-
-image = pipe(prompt_embeds=prompt_embeds, generator=generator, num_inference_steps=20).images[0]
-image
-```
-
-<div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/compel/forest_1.png"/>
-</div>
-
-To downweight parts of the prompt, use the `-` suffix:
-
-```py
-prompt = "a red------- cat playing with a ball"
-prompt_embeds = compel_proc(prompt)
-
-generator = torch.manual_seed(33)
-
-image = pipe(prompt_embeds=prompt_embeds, generator=generator, num_inference_steps=20).images[0]
-image
-```
-
-<div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-neg.png"/>
-</div>
-
-You can even up or downweight multiple concepts in the same prompt:
-
-```py
-prompt = "a red cat++ playing with a ball----"
-prompt_embeds = compel_proc(prompt)
-
-generator = torch.manual_seed(33)
-
-image = pipe(prompt_embeds=prompt_embeds, generator=generator, num_inference_steps=20).images[0]
-image
-```
-
-<div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-pos-neg.png"/>
-</div>
-
-### Blending
-
-You can also create a weighted *blend* of prompts by adding `.blend()` to a list of prompts and passing it some weights. Your blend may not always produce the result you expect because it breaks some assumptions about how the text encoder functions, so just have fun and experiment with it!
-
-```py
-prompt_embeds = compel_proc('("a red cat playing with a ball", "jungle").blend(0.7, 0.8)')
-generator = torch.Generator(device="cuda").manual_seed(33)
-
-image = pipe(prompt_embeds=prompt_embeds, generator=generator, num_inference_steps=20).images[0]
-image
-```
-
-<div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-blend.png"/>
-</div>
-
-### Conjunction
-
-A conjunction diffuses each prompt independently and concatenates their results by their weighted sum. Add `.and()` to the end of a list of prompts to create a conjunction:
-
-```py
-prompt_embeds = compel_proc('["a red cat", "playing with a", "ball"].and()')
-generator = torch.Generator(device="cuda").manual_seed(55)
-
-image = pipe(prompt_embeds=prompt_embeds, generator=generator, num_inference_steps=20).images[0]
-image
-```
-
-<div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-conj.png"/>
-</div>
+> [!TIP]
+> Refer to the [sd_embed](https://github.com/xhinker/sd_embed) repository for additional details about long prompt weighting for FLUX.1, Stable Cascade, and Stable Diffusion 1.5.
 
 ### Textual inversion
 
@@ -363,35 +326,63 @@ Create a pipeline and use the [`~loaders.TextualInversionLoaderMixin.load_textua
 ```py
 import torch
 from diffusers import StableDiffusionPipeline
-from compel import Compel, DiffusersTextualInversionManager
 
 pipe = StableDiffusionPipeline.from_pretrained(
-  "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16,
-  use_safetensors=True, variant="fp16").to("cuda")
+  "stable-diffusion-v1-5/stable-diffusion-v1-5",
+  torch_dtype=torch.float16,
+).to("cuda")
 pipe.load_textual_inversion("sd-concepts-library/midjourney-style")
 ```
 
-Compel provides a `DiffusersTextualInversionManager` class to simplify prompt weighting with textual inversion. Instantiate `DiffusersTextualInversionManager` and pass it to the `Compel` class:
+Add the `<midjourney-style>` text to the prompt to trigger the textual inversion.
 
 ```py
-textual_inversion_manager = DiffusersTextualInversionManager(pipe)
-compel_proc = Compel(
-    tokenizer=pipe.tokenizer,
-    text_encoder=pipe.text_encoder,
-    textual_inversion_manager=textual_inversion_manager)
+from sd_embed.embedding_funcs import get_weighted_text_embeddings_sd15
+
+prompt = """<midjourney-style> A whimsical and creative image depicting a hybrid creature that is a mix of a waffle and a hippopotamus. 
+This imaginative creature features the distinctive, bulky body of a hippo, 
+but with a texture and appearance resembling a golden-brown, crispy waffle. 
+The creature might have elements like waffle squares across its skin and a syrup-like sheen. 
+It's set in a surreal environment that playfully combines a natural water habitat of a hippo with elements of a breakfast table setting, 
+possibly including oversized utensils or plates in the background. 
+The image should evoke a sense of playful absurdity and culinary fantasy.
+"""
+
+neg_prompt = """\
+skin spots,acnes,skin blemishes,age spot,(ugly:1.2),(duplicate:1.2),(morbid:1.21),(mutilated:1.2),\
+(tranny:1.2),mutated hands,(poorly drawn hands:1.5),blurry,(bad anatomy:1.2),(bad proportions:1.3),\
+extra limbs,(disfigured:1.2),(missing arms:1.2),(extra legs:1.2),(fused fingers:1.5),\
+(too many fingers:1.5),(unclear eyes:1.2),lowers,bad hands,missing fingers,extra digit,\
+bad hands,missing fingers,(extra arms and legs),(worst quality:2),(low quality:2),\
+(normal quality:2),lowres,((monochrome)),((grayscale))
+"""
 ```
 
-Incorporate the concept to condition a prompt with using the `<concept>` syntax:
+Use the `get_weighted_text_embeddings_sd15` function to generate the prompt embeddings and the negative prompt embeddings.
 
 ```py
-prompt_embeds = compel_proc('("A red cat++ playing with a ball <midjourney-style>")')
+( 
+  prompt_embeds,
+  prompt_neg_embeds,
+) = get_weighted_text_embeddings_sd15(
+    pipe,
+    prompt=prompt,
+    neg_prompt=neg_prompt
+)
 
-image = pipe(prompt_embeds=prompt_embeds).images[0]
+image = pipe(
+    prompt_embeds=prompt_embeds,
+    negative_prompt_embeds=prompt_neg_embeds,
+    height=768,
+    width=896,
+    guidance_scale=4.0,
+    generator=torch.Generator("cuda").manual_seed(2)
+).images[0]
 image
 ```
 
 <div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-text-inversion.png"/>
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sd_embed_textual_inversion.png"/>
 </div>
 
 ### DreamBooth
@@ -401,70 +392,44 @@ image
 ```py
 import torch
 from diffusers import DiffusionPipeline, UniPCMultistepScheduler
-from compel import Compel
 
 pipe = DiffusionPipeline.from_pretrained("sd-dreambooth-library/dndcoverart-v1", torch_dtype=torch.float16).to("cuda")
 pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
 ```
 
-Create a `Compel` class with a tokenizer and text encoder, and pass your prompt to it. Depending on the model you use, you'll need to incorporate the model's unique identifier into your prompt. For example, the `dndcoverart-v1` model uses the identifier `dndcoverart`:
+Depending on the model you use, you'll need to incorporate the model's unique identifier into your prompt. For example, the `dndcoverart-v1` model uses the identifier `dndcoverart`:
 
 ```py
-compel_proc = Compel(tokenizer=pipe.tokenizer, text_encoder=pipe.text_encoder)
-prompt_embeds = compel_proc('("magazine cover of a dndcoverart dragon, high quality, intricate details, larry elmore art style").and()')
-image = pipe(prompt_embeds=prompt_embeds).images[0]
-image
-```
+from sd_embed.embedding_funcs import get_weighted_text_embeddings_sd15
 
-<div class="flex justify-center">
-  <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-dreambooth.png"/>
-</div>
+prompt = """dndcoverart of A whimsical and creative image depicting a hybrid creature that is a mix of a waffle and a hippopotamus. 
+This imaginative creature features the distinctive, bulky body of a hippo, 
+but with a texture and appearance resembling a golden-brown, crispy waffle. 
+The creature might have elements like waffle squares across its skin and a syrup-like sheen. 
+It's set in a surreal environment that playfully combines a natural water habitat of a hippo with elements of a breakfast table setting, 
+possibly including oversized utensils or plates in the background. 
+The image should evoke a sense of playful absurdity and culinary fantasy.
+"""
 
-### Stable Diffusion XL
+neg_prompt = """\
+skin spots,acnes,skin blemishes,age spot,(ugly:1.2),(duplicate:1.2),(morbid:1.21),(mutilated:1.2),\
+(tranny:1.2),mutated hands,(poorly drawn hands:1.5),blurry,(bad anatomy:1.2),(bad proportions:1.3),\
+extra limbs,(disfigured:1.2),(missing arms:1.2),(extra legs:1.2),(fused fingers:1.5),\
+(too many fingers:1.5),(unclear eyes:1.2),lowers,bad hands,missing fingers,extra digit,\
+bad hands,missing fingers,(extra arms and legs),(worst quality:2),(low quality:2),\
+(normal quality:2),lowres,((monochrome)),((grayscale))
+"""
 
-Stable Diffusion XL (SDXL) has two tokenizers and text encoders so it's usage is a bit different. To address this, you should pass both tokenizers and encoders to the `Compel` class:
-
-```py
-from compel import Compel, ReturnedEmbeddingsType
-from diffusers import DiffusionPipeline
-from diffusers.utils import make_image_grid
-import torch
-
-pipeline = DiffusionPipeline.from_pretrained(
-  "stabilityai/stable-diffusion-xl-base-1.0",
-  variant="fp16",
-  use_safetensors=True,
-  torch_dtype=torch.float16
-).to("cuda")
-
-compel = Compel(
-  tokenizer=[pipeline.tokenizer, pipeline.tokenizer_2] ,
-  text_encoder=[pipeline.text_encoder, pipeline.text_encoder_2],
-  returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
-  requires_pooled=[False, True]
+(
+    prompt_embeds
+    , prompt_neg_embeds
+) = get_weighted_text_embeddings_sd15(
+    pipe
+    , prompt = prompt
+    , neg_prompt = neg_prompt
 )
 ```
 
-This time, let's upweight "ball" by a factor of 1.5 for the first prompt, and downweight "ball" by 0.6 for the second prompt. The [`StableDiffusionXLPipeline`] also requires [`pooled_prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLInpaintPipeline.__call__.pooled_prompt_embeds) (and optionally [`negative_pooled_prompt_embeds`](https://huggingface.co/docs/diffusers/en/api/pipelines/stable_diffusion/stable_diffusion_xl#diffusers.StableDiffusionXLInpaintPipeline.__call__.negative_pooled_prompt_embeds)) so you should pass those to the pipeline along with the conditioning tensors:
-
-```py
-# apply weights
-prompt = ["a red cat playing with a (ball)1.5", "a red cat playing with a (ball)0.6"]
-conditioning, pooled = compel(prompt)
-
-# generate image
-generator = [torch.Generator().manual_seed(33) for _ in range(len(prompt))]
-images = pipeline(prompt_embeds=conditioning, pooled_prompt_embeds=pooled, generator=generator, num_inference_steps=30).images
-make_image_grid(images, rows=1, cols=2)
-```
-
-<div class="flex gap-4">
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/compel/sdxl_ball1.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">"a red cat playing with a (ball)1.5"</figcaption>
-  </div>
-  <div>
-    <img class="rounded-xl" src="https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/compel/sdxl_ball2.png"/>
-    <figcaption class="mt-2 text-center text-sm text-gray-500">"a red cat playing with a (ball)0.6"</figcaption>
-  </div>
+<div class="flex justify-center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/sd_embed_dreambooth.png"/>
 </div>

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -227,7 +227,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     autocast_ctx = nullcontext()
 
     with autocast_ctx:
@@ -880,9 +880,7 @@ class TokenEmbeddingsHandler:
         idx_to_text_encoder_name = {0: "clip_l", 1: "t5"}
         for idx, text_encoder in enumerate(self.text_encoders):
             train_ids = self.train_ids if idx == 0 else self.train_ids_t5
-            embeds = (
-                text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.encoder.embed_tokens
-            )
+            embeds = text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.shared
             assert embeds.weight.data.shape[0] == len(self.tokenizers[idx]), "Tokenizers should be the same."
             new_token_embeddings = embeds.weight.data[train_ids]
 
@@ -904,9 +902,7 @@ class TokenEmbeddingsHandler:
     @torch.no_grad()
     def retract_embeddings(self):
         for idx, text_encoder in enumerate(self.text_encoders):
-            embeds = (
-                text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.encoder.embed_tokens
-            )
+            embeds = text_encoder.text_model.embeddings.token_embedding if idx == 0 else text_encoder.shared
             index_no_updates = self.embeddings_settings[f"index_no_updates_{idx}"]
             embeds.weight.data[index_no_updates] = (
                 self.embeddings_settings[f"original_embeddings_{idx}"][index_no_updates]
@@ -1749,7 +1745,7 @@ def main(args):
         if args.enable_t5_ti:  # whether to do pivotal tuning/textual inversion for T5 as well
             text_lora_parameters_two = []
             for name, param in text_encoder_two.named_parameters():
-                if "token_embedding" in name:
+                if "shared" in name:
                     # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
                     param.data = param.to(dtype=torch.float32)
                     param.requires_grad = True

examples/advanced_diffusion_training/train_dreambooth_lora_sd15_advanced.py

@@ -1883,7 +1883,11 @@ def main(args):
                 pipeline.set_progress_bar_config(disable=True)
 
                 # run inference
-                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+                generator = (
+                    torch.Generator(device=accelerator.device).manual_seed(args.seed)
+                    if args.seed is not None
+                    else None
+                )
                 pipeline_args = {"prompt": args.validation_prompt}
 
                 if torch.backends.mps.is_available():
@@ -1987,7 +1991,9 @@ def main(args):
                 )
             # run inference
             pipeline = pipeline.to(accelerator.device)
-            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+            generator = (
+                torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
+            )
             images = [
                 pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
                 for _ in range(args.num_validation_images)

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -269,7 +269,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     # Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
     # way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
     if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:

examples/cogvideo/train_cogvideox_image_to_video_lora.py

@@ -722,7 +722,7 @@ def log_validation(
     # pipe.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
 
     videos = []
     for _ in range(args.num_validation_videos):

examples/cogvideo/train_cogvideox_lora.py

@@ -739,7 +739,7 @@ def log_validation(
     # pipe.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
 
     videos = []
     for _ in range(args.num_validation_videos):

examples/community/README.md

@@ -10,6 +10,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 
 | Example                                                                                                                               | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              | Code Example                                                                              | Colab                                                                                                                                                                                                              |                                                        Author |
 |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
+|Spatiotemporal Skip Guidance (STG)|[Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling](https://arxiv.org/abs/2411.18664) (CVPR 2025) enhances video diffusion models by generating a weaker model through layer skipping and using it as guidance, improving fidelity in models like HunyuanVideo, LTXVideo, and Mochi.|[Spatiotemporal Skip Guidance](#spatiotemporal-skip-guidance)|-|[Junha Hyung](https://junhahyung.github.io/), [Kinam Kim](https://kinam0252.github.io/)|
 |Adaptive Mask Inpainting|Adaptive Mask Inpainting algorithm from [Beyond the Contact: Discovering Comprehensive Affordance for 3D Objects from Pre-trained 2D Diffusion Models](https://github.com/snuvclab/coma) (ECCV '24, Oral) provides a way to insert human inside the scene image without altering the background, by inpainting with adapting mask.|[Adaptive Mask Inpainting](#adaptive-mask-inpainting)|-|[Hyeonwoo Kim](https://sshowbiz.xyz),[Sookwan Han](https://jellyheadandrew.github.io)|
 |Flux with CFG|[Flux with CFG](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md) provides an implementation of using CFG in [Flux](https://blackforestlabs.ai/announcing-black-forest-labs/).|[Flux with CFG](#flux-with-cfg)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/flux_with_cfg.ipynb)|[Linoy Tsaban](https://github.com/linoytsaban), [Apolinário](https://github.com/apolinario), and [Sayak Paul](https://github.com/sayakpaul)|
 |Differential Diffusion|[Differential Diffusion](https://github.com/exx8/differential-diffusion) modifies an image according to a text prompt, and according to a map that specifies the amount of change in each region.|[Differential Diffusion](#differential-diffusion)|[![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/exx8/differential-diffusion) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/exx8/differential-diffusion/blob/main/examples/SD2.ipynb)|[Eran Levin](https://github.com/exx8) and [Ohad Fried](https://www.ohadf.com/)|
@@ -53,6 +54,7 @@ Please also check out our [Community Scripts](https://github.com/huggingface/dif
 | Stable Diffusion Mixture Tiling Pipeline SD 1.5 | A pipeline generates cohesive images by integrating multiple diffusion processes, each focused on a specific image region and considering boundary effects for smooth blending | [Stable Diffusion Mixture Tiling Pipeline SD 1.5](#stable-diffusion-mixture-tiling-pipeline-sd-15) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/albarji/mixture-of-diffusers) | [Álvaro B Jiménez](https://github.com/albarji/) |
 | Stable Diffusion Mixture Canvas Pipeline SD 1.5 | A pipeline generates cohesive images by integrating multiple diffusion processes, each focused on a specific image region and considering boundary effects for smooth blending. Works by defining a list of Text2Image region objects that detail the region of influence of each diffuser. | [Stable Diffusion Mixture Canvas Pipeline SD 1.5](#stable-diffusion-mixture-canvas-pipeline-sd-15) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/albarji/mixture-of-diffusers) | [Álvaro B Jiménez](https://github.com/albarji/) |
 | Stable Diffusion Mixture Tiling Pipeline SDXL | A pipeline generates cohesive images by integrating multiple diffusion processes, each focused on a specific image region and considering boundary effects for smooth blending | [Stable Diffusion Mixture Tiling Pipeline SDXL](#stable-diffusion-mixture-tiling-pipeline-sdxl) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/elismasilva/mixture-of-diffusers-sdxl-tiling) | [Eliseu Silva](https://github.com/DEVAIEXP/) |
+| Stable Diffusion MoD ControlNet Tile SR Pipeline SDXL | This is an advanced pipeline that leverages ControlNet Tile and Mixture-of-Diffusers techniques, integrating tile diffusion directly into the latent space denoising process. Designed to overcome the limitations of conventional pixel-space tile processing, this pipeline delivers Super Resolution (SR) upscaling for higher-quality images, reduced processing time, and greater adaptability. | [Stable Diffusion MoD ControlNet Tile SR Pipeline SDXL](#stable-diffusion-mod-controlnet-tile-sr-pipeline-sdxl) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/elismasilva/mod-control-tile-upscaler-sdxl) | [Eliseu Silva](https://github.com/DEVAIEXP/) |
 | FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | [Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/stable_diffusion_fabric.ipynb)| [Shauray Singh](https://shauray8.github.io/about_shauray/) |
 | sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
 | sketch inpaint xl - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion XL Pipeline](#stable-diffusion-xl-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
@@ -82,6 +84,7 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | [🪆Matryoshka Diffusion Models](https://huggingface.co/papers/2310.15111) | A diffusion process that denoises inputs at multiple resolutions jointly and uses a NestedUNet architecture where features and parameters for small scale inputs are nested within those of the large scales. See [original codebase](https://github.com/apple/ml-mdm). | [🪆Matryoshka Diffusion Models](#matryoshka-diffusion-models) | [![Hugging Face Space](https://img.shields.io/badge/🤗%20Hugging%20Face-Space-yellow)](https://huggingface.co/spaces/pcuenq/mdm) [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/gist/tolgacangoz/1f54875fc7aeaabcf284ebde64820966/matryoshka_hf.ipynb) | [M. Tolga Cangöz](https://github.com/tolgacangoz) |
 | Stable Diffusion XL Attentive Eraser Pipeline |[[AAAI2025 Oral] Attentive Eraser](https://github.com/Anonym0u3/AttentiveEraser) is a novel tuning-free method that enhances object removal capabilities in pre-trained diffusion models.|[Stable Diffusion XL Attentive Eraser Pipeline](#stable-diffusion-xl-attentive-eraser-pipeline)|-|[Wenhao Sun](https://github.com/Anonym0u3) and [Benlei Cui](https://github.com/Benny079)|
 | Perturbed-Attention Guidance |StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).|[Perturbed-Attention Guidance](#perturbed-attention-guidance)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/perturbed_attention_guidance.ipynb)|[Hyoungwon Cho](https://github.com/HyoungwonCho)|
+| CogVideoX DDIM Inversion Pipeline | Implementation of DDIM inversion and guided attention-based editing denoising process on CogVideoX. | [CogVideoX DDIM Inversion Pipeline](#cogvideox-ddim-inversion-pipeline) | - | [LittleNyima](https://github.com/LittleNyima) |
 
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
 
@@ -91,6 +94,55 @@ pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion
 
 ## Example usages
 
+### Spatiotemporal Skip Guidance
+
+**Junha Hyung\*, Kinam Kim\*, Susung Hong, Min-Jung Kim, Jaegul Choo**
+
+**KAIST AI, University of Washington**
+
+[*Spatiotemporal Skip Guidance (STG) for Enhanced Video Diffusion Sampling*](https://arxiv.org/abs/2411.18664) (CVPR 2025) is a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree.
+
+Following is the example video of STG applied to Mochi.
+
+
+https://github.com/user-attachments/assets/148adb59-da61-4c50-9dfa-425dcb5c23b3
+
+More examples and information can be found on the [GitHub repository](https://github.com/junhahyung/STGuidance) and the [Project website](https://junhahyung.github.io/STGuidance/).
+
+#### Usage example
+```python
+import torch
+from pipeline_stg_mochi import MochiSTGPipeline
+from diffusers.utils import export_to_video
+
+# Load the pipeline
+pipe = MochiSTGPipeline.from_pretrained("genmo/mochi-1-preview", variant="bf16", torch_dtype=torch.bfloat16)
+
+# Enable memory savings
+pipe = pipe.to("cuda")
+
+#--------Option--------#
+prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
+stg_applied_layers_idx = [34]
+stg_mode = "STG"
+stg_scale = 1.0 # 0.0 for CFG
+#----------------------#
+
+# Generate video frames
+frames = pipe(
+    prompt, 
+    height=480,
+    width=480,
+    num_frames=81,
+    stg_applied_layers_idx=stg_applied_layers_idx,
+    stg_scale=stg_scale,
+    generator = torch.Generator().manual_seed(42),
+    do_rescaling=do_rescaling,
+).frames[0]
+
+export_to_video(frames, "output.mp4", fps=30)
+```
+
 ### Adaptive Mask Inpainting
 
 **Hyeonwoo Kim\*, Sookwan Han\*, Patrick Kwon, Hanbyul Joo**
@@ -2630,6 +2682,103 @@ image = pipe(
 
 ![mixture_tiling_results](https://huggingface.co/datasets/elismasilva/results/resolve/main/mixture_of_diffusers_sdxl_1.png)
 
+### Stable Diffusion MoD ControlNet Tile SR Pipeline SDXL
+
+This pipeline implements the [MoD (Mixture-of-Diffusers)]("https://arxiv.org/pdf/2408.06072") tiled diffusion technique and combines it with SDXL's ControlNet Tile process to generate SR images.
+
+This works better with 4x scales, but you can try adjusts parameters to higher scales.
+
+````python
+import torch
+from diffusers import DiffusionPipeline, ControlNetUnionModel, AutoencoderKL, UniPCMultistepScheduler, UNet2DConditionModel
+from diffusers.utils import load_image
+from PIL import Image
+
+device = "cuda"
+
+# Initialize the models and pipeline
+controlnet = ControlNetUnionModel.from_pretrained(
+    "brad-twinkl/controlnet-union-sdxl-1.0-promax", torch_dtype=torch.float16
+).to(device=device)
+vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to(device=device)
+
+model_id = "SG161222/RealVisXL_V5.0"
+pipe = DiffusionPipeline.from_pretrained(
+    model_id,
+    torch_dtype=torch.float16,
+    vae=vae,
+    controlnet=controlnet,
+    custom_pipeline="mod_controlnet_tile_sr_sdxl",    
+    use_safetensors=True,
+    variant="fp16",
+).to(device)
+
+unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
+
+#pipe.enable_model_cpu_offload()  # << Enable this if you have limited VRAM
+pipe.enable_vae_tiling() # << Enable this if you have limited VRAM
+pipe.enable_vae_slicing() # << Enable this if you have limited VRAM
+
+# Set selected scheduler
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+# Load image
+control_image = load_image("https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/1.jpg")
+original_height = control_image.height
+original_width = control_image.width
+print(f"Current resolution: H:{original_height} x W:{original_width}")
+
+# Pre-upscale image for tiling
+resolution = 4096
+tile_gaussian_sigma = 0.3
+max_tile_size = 1024 # or 1280
+
+current_size = max(control_image.size)
+scale_factor = max(2, resolution / current_size)
+new_size = (int(control_image.width * scale_factor), int(control_image.height * scale_factor))
+image = control_image.resize(new_size, Image.LANCZOS)
+
+# Update target height and width
+target_height = image.height
+target_width = image.width
+print(f"Target resolution: H:{target_height} x W:{target_width}")
+
+# Calculate overlap size
+normal_tile_overlap, border_tile_overlap = pipe.calculate_overlap(target_width, target_height)
+
+# Set other params
+tile_weighting_method = pipe.TileWeightingMethod.COSINE.value
+guidance_scale = 4
+num_inference_steps = 35
+denoising_strenght = 0.65
+controlnet_strength = 1.0
+prompt = "high-quality, noise-free edges, high quality, 4k, hd, 8k"
+negative_prompt = "blurry, pixelated, noisy, low resolution, artifacts, poor details"
+
+# Image generation
+generated_image = pipe(
+    image=image,
+    control_image=control_image,
+    control_mode=[6],
+    controlnet_conditioning_scale=float(controlnet_strength),
+    prompt=prompt,
+    negative_prompt=negative_prompt,
+    normal_tile_overlap=normal_tile_overlap,
+    border_tile_overlap=border_tile_overlap,
+    height=target_height,
+    width=target_width,
+    original_size=(original_width, original_height),
+    target_size=(target_width, target_height),
+    guidance_scale=guidance_scale,        
+    strength=float(denoising_strenght),
+    tile_weighting_method=tile_weighting_method,
+    max_tile_size=max_tile_size,
+    tile_gaussian_sigma=float(tile_gaussian_sigma),
+    num_inference_steps=num_inference_steps,
+)["images"][0]
+````
+![Upscaled](https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/1_input_4x.png)
+
 ### TensorRT Inpainting Stable Diffusion Pipeline
 
 The TensorRT Pipeline can be used to accelerate the Inpainting Stable Diffusion Inference run.
@@ -5124,3 +5273,39 @@ with torch.no_grad():
 
 In the folder examples/pixart there is also a script that can be used to train new models.
 Please check the script `train_controlnet_hf_diffusers.sh` on how to start the training.
+
+# CogVideoX DDIM Inversion Pipeline
+
+This implementation performs DDIM inversion on the video based on CogVideoX and uses guided attention to reconstruct or edit the inversion latents.
+
+## Example Usage
+
+```python
+import torch
+
+from examples.community.cogvideox_ddim_inversion import CogVideoXPipelineForDDIMInversion
+
+
+# Load pretrained pipeline
+pipeline = CogVideoXPipelineForDDIMInversion.from_pretrained(
+    "THUDM/CogVideoX1.5-5B",
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+
+# Run DDIM inversion, and the videos will be generated in the output_path
+output = pipeline_for_inversion(
+    prompt="prompt that describes the edited video",
+    video_path="path/to/input.mp4",
+    guidance_scale=6.0,
+    num_inference_steps=50,
+    skip_frames_start=0,
+    skip_frames_end=0,
+    frame_sample_step=None,
+    max_num_frames=81,
+    width=720,
+    height=480,
+    seed=42,
+)
+pipeline.export_latents_to_video(output.inverse_latents[-1], "path/to/inverse_video.mp4", fps=8)
+pipeline.export_latents_to_video(output.recon_latents[-1], "path/to/recon_video.mp4", fps=8)
+```

examples/community/checkpoint_merger.py

@@ -92,9 +92,13 @@ class CheckpointMergerPipeline(DiffusionPipeline):
         token = kwargs.pop("token", None)
         variant = kwargs.pop("variant", None)
         revision = kwargs.pop("revision", None)
-        torch_dtype = kwargs.pop("torch_dtype", None)
+        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
         device_map = kwargs.pop("device_map", None)
 
+        if not isinstance(torch_dtype, torch.dtype):
+            torch_dtype = torch.float32
+            print(f"Passed `torch_dtype` {torch_dtype} is not a `torch.dtype`. Defaulting to `torch.float32`.")
+
         alpha = kwargs.pop("alpha", 0.5)
         interp = kwargs.pop("interp", None)
 

examples/community/cogvideox_ddim_inversion.py

@@ -0,0 +1,645 @@
+"""
+This script performs DDIM inversion for video frames using a pre-trained model and generates
+a video reconstruction based on a provided prompt. It utilizes the CogVideoX pipeline to
+process video frames, apply the DDIM inverse scheduler, and produce an output video.
+
+**Please notice that this script is based on the CogVideoX 5B model, and would not generate
+a good result for 2B variants.**
+
+Usage:
+    python cogvideox_ddim_inversion.py
+        --model-path /path/to/model
+        --prompt "a prompt"
+        --video-path /path/to/video.mp4
+        --output-path /path/to/output
+
+For more details about the cli arguments, please run `python cogvideox_ddim_inversion.py --help`.
+
+Author:
+    LittleNyima <littlenyima[at]163[dot]com>
+"""
+
+import argparse
+import math
+import os
+from typing import Any, Dict, List, Optional, Tuple, TypedDict, Union, cast
+
+import torch
+import torch.nn.functional as F
+import torchvision.transforms as T
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.models.attention_processor import Attention, CogVideoXAttnProcessor2_0
+from diffusers.models.autoencoders import AutoencoderKLCogVideoX
+from diffusers.models.embeddings import apply_rotary_emb
+from diffusers.models.transformers.cogvideox_transformer_3d import CogVideoXBlock, CogVideoXTransformer3DModel
+from diffusers.pipelines.cogvideo.pipeline_cogvideox import CogVideoXPipeline, retrieve_timesteps
+from diffusers.schedulers import CogVideoXDDIMScheduler, DDIMInverseScheduler
+from diffusers.utils import export_to_video
+
+
+# Must import after torch because this can sometimes lead to a nasty segmentation fault, or stack smashing error.
+# Very few bug reports but it happens. Look in decord Github issues for more relevant information.
+import decord  # isort: skip
+
+
+class DDIMInversionArguments(TypedDict):
+    model_path: str
+    prompt: str
+    video_path: str
+    output_path: str
+    guidance_scale: float
+    num_inference_steps: int
+    skip_frames_start: int
+    skip_frames_end: int
+    frame_sample_step: Optional[int]
+    max_num_frames: int
+    width: int
+    height: int
+    fps: int
+    dtype: torch.dtype
+    seed: int
+    device: torch.device
+
+
+def get_args() -> DDIMInversionArguments:
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("--model_path", type=str, required=True, help="Path of the pretrained model")
+    parser.add_argument("--prompt", type=str, required=True, help="Prompt for the direct sample procedure")
+    parser.add_argument("--video_path", type=str, required=True, help="Path of the video for inversion")
+    parser.add_argument("--output_path", type=str, default="output", help="Path of the output videos")
+    parser.add_argument("--guidance_scale", type=float, default=6.0, help="Classifier-free guidance scale")
+    parser.add_argument("--num_inference_steps", type=int, default=50, help="Number of inference steps")
+    parser.add_argument("--skip_frames_start", type=int, default=0, help="Number of skipped frames from the start")
+    parser.add_argument("--skip_frames_end", type=int, default=0, help="Number of skipped frames from the end")
+    parser.add_argument("--frame_sample_step", type=int, default=None, help="Temporal stride of the sampled frames")
+    parser.add_argument("--max_num_frames", type=int, default=81, help="Max number of sampled frames")
+    parser.add_argument("--width", type=int, default=720, help="Resized width of the video frames")
+    parser.add_argument("--height", type=int, default=480, help="Resized height of the video frames")
+    parser.add_argument("--fps", type=int, default=8, help="Frame rate of the output videos")
+    parser.add_argument("--dtype", type=str, default="bf16", choices=["bf16", "fp16"], help="Dtype of the model")
+    parser.add_argument("--seed", type=int, default=42, help="Seed for the random number generator")
+    parser.add_argument("--device", type=str, default="cuda", choices=["cuda", "cpu"], help="Device for inference")
+
+    args = parser.parse_args()
+    args.dtype = torch.bfloat16 if args.dtype == "bf16" else torch.float16
+    args.device = torch.device(args.device)
+
+    return DDIMInversionArguments(**vars(args))
+
+
+class CogVideoXAttnProcessor2_0ForDDIMInversion(CogVideoXAttnProcessor2_0):
+    def __init__(self):
+        super().__init__()
+
+    def calculate_attention(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn: Attention,
+        batch_size: int,
+        image_seq_length: int,
+        text_seq_length: int,
+        attention_mask: Optional[torch.Tensor],
+        image_rotary_emb: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        r"""
+        Core attention computation with inversion-guided RoPE integration.
+
+        Args:
+            query (`torch.Tensor`): `[batch_size, seq_len, dim]` query tensor
+            key (`torch.Tensor`): `[batch_size, seq_len, dim]` key tensor
+            value (`torch.Tensor`): `[batch_size, seq_len, dim]` value tensor
+            attn (`Attention`): Parent attention module with projection layers
+            batch_size (`int`): Effective batch size (after chunk splitting)
+            image_seq_length (`int`): Length of image feature sequence
+            text_seq_length (`int`): Length of text feature sequence
+            attention_mask (`Optional[torch.Tensor]`): Attention mask tensor
+            image_rotary_emb (`Optional[torch.Tensor]`): Rotary embeddings for image positions
+
+        Returns:
+            `Tuple[torch.Tensor, torch.Tensor]`:
+                (1) hidden_states: [batch_size, image_seq_length, dim] processed image features
+                (2) encoder_hidden_states: [batch_size, text_seq_length, dim] processed text features
+        """
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
+            if not attn.is_cross_attention:
+                if key.size(2) == query.size(2):  # Attention for reference hidden states
+                    key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
+                else:  # RoPE should be applied to each group of image tokens
+                    key[:, :, text_seq_length : text_seq_length + image_seq_length] = apply_rotary_emb(
+                        key[:, :, text_seq_length : text_seq_length + image_seq_length], image_rotary_emb
+                    )
+                    key[:, :, text_seq_length * 2 + image_seq_length :] = apply_rotary_emb(
+                        key[:, :, text_seq_length * 2 + image_seq_length :], image_rotary_emb
+                    )
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+        return hidden_states, encoder_hidden_states
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        r"""
+        Process the dual-path attention for the inversion-guided denoising procedure.
+
+        Args:
+            attn (`Attention`): Parent attention module
+            hidden_states (`torch.Tensor`): `[batch_size, image_seq_len, dim]` Image tokens
+            encoder_hidden_states (`torch.Tensor`): `[batch_size, text_seq_len, dim]` Text tokens
+            attention_mask (`Optional[torch.Tensor]`): Optional attention mask
+            image_rotary_emb (`Optional[torch.Tensor]`): Rotary embeddings for image tokens
+
+        Returns:
+            `Tuple[torch.Tensor, torch.Tensor]`:
+                (1) Final hidden states: `[batch_size, image_seq_length, dim]` Resulting image tokens
+                (2) Final encoder states: `[batch_size, text_seq_length, dim]` Resulting text tokens
+        """
+        image_seq_length = hidden_states.size(1)
+        text_seq_length = encoder_hidden_states.size(1)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query, query_reference = query.chunk(2)
+        key, key_reference = key.chunk(2)
+        value, value_reference = value.chunk(2)
+        batch_size = batch_size // 2
+
+        hidden_states, encoder_hidden_states = self.calculate_attention(
+            query=query,
+            key=torch.cat((key, key_reference), dim=1),
+            value=torch.cat((value, value_reference), dim=1),
+            attn=attn,
+            batch_size=batch_size,
+            image_seq_length=image_seq_length,
+            text_seq_length=text_seq_length,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+        hidden_states_reference, encoder_hidden_states_reference = self.calculate_attention(
+            query=query_reference,
+            key=key_reference,
+            value=value_reference,
+            attn=attn,
+            batch_size=batch_size,
+            image_seq_length=image_seq_length,
+            text_seq_length=text_seq_length,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+
+        return (
+            torch.cat((hidden_states, hidden_states_reference)),
+            torch.cat((encoder_hidden_states, encoder_hidden_states_reference)),
+        )
+
+
+class OverrideAttnProcessors:
+    r"""
+    Context manager for temporarily overriding attention processors in CogVideo transformer blocks.
+
+    Designed for DDIM inversion process, replaces original attention processors with
+    `CogVideoXAttnProcessor2_0ForDDIMInversion` and restores them upon exit. Uses Python context manager
+    pattern to safely manage processor replacement.
+
+    Typical usage:
+    ```python
+    with OverrideAttnProcessors(transformer):
+        # Perform DDIM inversion operations
+    ```
+
+    Args:
+        transformer (`CogVideoXTransformer3DModel`):
+            The transformer model containing attention blocks to be modified. Should have
+            `transformer_blocks` attribute containing `CogVideoXBlock` instances.
+    """
+
+    def __init__(self, transformer: CogVideoXTransformer3DModel):
+        self.transformer = transformer
+        self.original_processors = {}
+
+    def __enter__(self):
+        for block in self.transformer.transformer_blocks:
+            block = cast(CogVideoXBlock, block)
+            self.original_processors[id(block)] = block.attn1.get_processor()
+            block.attn1.set_processor(CogVideoXAttnProcessor2_0ForDDIMInversion())
+
+    def __exit__(self, _0, _1, _2):
+        for block in self.transformer.transformer_blocks:
+            block = cast(CogVideoXBlock, block)
+            block.attn1.set_processor(self.original_processors[id(block)])
+
+
+def get_video_frames(
+    video_path: str,
+    width: int,
+    height: int,
+    skip_frames_start: int,
+    skip_frames_end: int,
+    max_num_frames: int,
+    frame_sample_step: Optional[int],
+) -> torch.FloatTensor:
+    """
+    Extract and preprocess video frames from a video file for VAE processing.
+
+    Args:
+        video_path (`str`): Path to input video file
+        width (`int`): Target frame width for decoding
+        height (`int`): Target frame height for decoding
+        skip_frames_start (`int`): Number of frames to skip at video start
+        skip_frames_end (`int`): Number of frames to skip at video end
+        max_num_frames (`int`): Maximum allowed number of output frames
+        frame_sample_step (`Optional[int]`):
+            Frame sampling step size. If None, automatically calculated as:
+            (total_frames - skipped_frames) // max_num_frames
+
+    Returns:
+        `torch.FloatTensor`: Preprocessed frames in `[F, C, H, W]` format where:
+        - `F`: Number of frames (adjusted to 4k + 1 for VAE compatibility)
+        - `C`: Channels (3 for RGB)
+        - `H`: Frame height
+        - `W`: Frame width
+    """
+    with decord.bridge.use_torch():
+        video_reader = decord.VideoReader(uri=video_path, width=width, height=height)
+        video_num_frames = len(video_reader)
+        start_frame = min(skip_frames_start, video_num_frames)
+        end_frame = max(0, video_num_frames - skip_frames_end)
+
+        if end_frame <= start_frame:
+            indices = [start_frame]
+        elif end_frame - start_frame <= max_num_frames:
+            indices = list(range(start_frame, end_frame))
+        else:
+            step = frame_sample_step or (end_frame - start_frame) // max_num_frames
+            indices = list(range(start_frame, end_frame, step))
+
+        frames = video_reader.get_batch(indices=indices)
+        frames = frames[:max_num_frames].float()  # ensure that we don't go over the limit
+
+        # Choose first (4k + 1) frames as this is how many is required by the VAE
+        selected_num_frames = frames.size(0)
+        remainder = (3 + selected_num_frames) % 4
+        if remainder != 0:
+            frames = frames[:-remainder]
+        assert frames.size(0) % 4 == 1
+
+        # Normalize the frames
+        transform = T.Lambda(lambda x: x / 255.0 * 2.0 - 1.0)
+        frames = torch.stack(tuple(map(transform, frames)), dim=0)
+
+        return frames.permute(0, 3, 1, 2).contiguous()  # [F, C, H, W]
+
+
+class CogVideoXDDIMInversionOutput:
+    inverse_latents: torch.FloatTensor
+    recon_latents: torch.FloatTensor
+
+    def __init__(self, inverse_latents: torch.FloatTensor, recon_latents: torch.FloatTensor):
+        self.inverse_latents = inverse_latents
+        self.recon_latents = recon_latents
+
+
+class CogVideoXPipelineForDDIMInversion(CogVideoXPipeline):
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: CogVideoXDDIMScheduler,
+    ):
+        super().__init__(
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            vae=vae,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        self.inverse_scheduler = DDIMInverseScheduler(**scheduler.config)
+
+    def encode_video_frames(self, video_frames: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        Encode video frames into latent space using Variational Autoencoder.
+
+        Args:
+            video_frames (`torch.FloatTensor`):
+                Input frames tensor in `[F, C, H, W]` format from `get_video_frames()`
+
+        Returns:
+            `torch.FloatTensor`: Encoded latents in `[1, F, D, H_latent, W_latent]` format where:
+            - `F`: Number of frames (same as input)
+            - `D`: Latent channel dimension
+            - `H_latent`: Latent space height (H // 2^vae.downscale_factor)
+            - `W_latent`: Latent space width (W // 2^vae.downscale_factor)
+        """
+        vae: AutoencoderKLCogVideoX = self.vae
+        video_frames = video_frames.to(device=vae.device, dtype=vae.dtype)
+        video_frames = video_frames.unsqueeze(0).permute(0, 2, 1, 3, 4)  # [B, C, F, H, W]
+        latent_dist = vae.encode(x=video_frames).latent_dist.sample().transpose(1, 2)
+        return latent_dist * vae.config.scaling_factor
+
+    @torch.no_grad()
+    def export_latents_to_video(self, latents: torch.FloatTensor, video_path: str, fps: int):
+        r"""
+        Decode latent vectors into video and export as video file.
+
+        Args:
+            latents (`torch.FloatTensor`): Encoded latents in `[B, F, D, H_latent, W_latent]` format from
+                `encode_video_frames()`
+            video_path (`str`): Output path for video file
+            fps (`int`): Target frames per second for output video
+        """
+        video = self.decode_latents(latents)
+        frames = self.video_processor.postprocess_video(video=video, output_type="pil")
+        os.makedirs(os.path.dirname(video_path), exist_ok=True)
+        export_to_video(video_frames=frames[0], output_video_path=video_path, fps=fps)
+
+    # Modified from CogVideoXPipeline.__call__
+    @torch.no_grad()
+    def sample(
+        self,
+        latents: torch.FloatTensor,
+        scheduler: Union[DDIMInverseScheduler, CogVideoXDDIMScheduler],
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        reference_latents: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Execute the core sampling loop for video generation/inversion using CogVideoX.
+
+        Implements the full denoising trajectory recording for both DDIM inversion and
+        generation processes. Supports dynamic classifier-free guidance and reference
+        latent conditioning.
+
+        Args:
+            latents (`torch.FloatTensor`):
+                Initial noise tensor of shape `[B, F, C, H, W]`.
+            scheduler (`Union[DDIMInverseScheduler, CogVideoXDDIMScheduler]`):
+                Scheduling strategy for diffusion process. Use:
+                (1) `DDIMInverseScheduler` for inversion
+                (2) `CogVideoXDDIMScheduler` for generation
+            prompt (`Optional[Union[str, List[str]]]`):
+                Text prompt(s) for conditional generation. Defaults to unconditional.
+            negative_prompt (`Optional[Union[str, List[str]]]`):
+                Negative prompt(s) for guidance. Requires `guidance_scale > 1`.
+            num_inference_steps (`int`):
+                Number of denoising steps. Affects quality/compute trade-off.
+            guidance_scale (`float`):
+                Classifier-free guidance weight. 1.0 = no guidance.
+            use_dynamic_cfg (`bool`):
+                Enable time-varying guidance scale (cosine schedule)
+            eta (`float`):
+                DDIM variance parameter (0 = deterministic process)
+            generator (`Optional[Union[torch.Generator, List[torch.Generator]]]`):
+                Random number generator(s) for reproducibility
+            attention_kwargs (`Optional[Dict[str, Any]]`):
+                Custom parameters for attention modules
+            reference_latents (`torch.FloatTensor`):
+                Reference latent trajectory for conditional sampling. Shape should match
+                `[T, B, F, C, H, W]` where `T` is number of timesteps
+
+        Returns:
+            `torch.FloatTensor`:
+                Full denoising trajectory tensor of shape `[T, B, F, C, H, W]`.
+        """
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            device=device,
+        )
+        if do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+        if reference_latents is not None:
+            prompt_embeds = torch.cat([prompt_embeds] * 2, dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(scheduler, num_inference_steps, device)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents.
+        latents = latents.to(device=device) * scheduler.init_noise_sigma
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+        if isinstance(scheduler, DDIMInverseScheduler):  # Inverse scheduler does not accept extra kwargs
+            extra_step_kwargs = {}
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(
+                height=latents.size(3) * self.vae_scale_factor_spatial,
+                width=latents.size(4) * self.vae_scale_factor_spatial,
+                num_frames=latents.size(1),
+                device=device,
+            )
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * scheduler.order, 0)
+
+        trajectory = torch.zeros_like(latents).unsqueeze(0).repeat(len(timesteps), 1, 1, 1, 1, 1)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+                if reference_latents is not None:
+                    reference = reference_latents[i]
+                    reference = torch.cat([reference] * 2) if do_classifier_free_guidance else reference
+                    latent_model_input = torch.cat([latent_model_input, reference], dim=0)
+                latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if reference_latents is not None:  # Recover the original batch size
+                    noise_pred, _ = noise_pred.chunk(2)
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the noisy sample x_t-1 -> x_t
+                latents = scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                latents = latents.to(prompt_embeds.dtype)
+                trajectory[i] = latents
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % scheduler.order == 0):
+                    progress_bar.update()
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        return trajectory
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: str,
+        video_path: str,
+        guidance_scale: float,
+        num_inference_steps: int,
+        skip_frames_start: int,
+        skip_frames_end: int,
+        frame_sample_step: Optional[int],
+        max_num_frames: int,
+        width: int,
+        height: int,
+        seed: int,
+    ):
+        """
+        Performs DDIM inversion on a video to reconstruct it with a new prompt.
+
+        Args:
+            prompt (`str`): The text prompt to guide the reconstruction.
+            video_path (`str`): Path to the input video file.
+            guidance_scale (`float`): Scale for classifier-free guidance.
+            num_inference_steps (`int`): Number of denoising steps.
+            skip_frames_start (`int`): Number of frames to skip from the beginning of the video.
+            skip_frames_end (`int`): Number of frames to skip from the end of the video.
+            frame_sample_step (`Optional[int]`): Step size for sampling frames. If None, all frames are used.
+            max_num_frames (`int`): Maximum number of frames to process.
+            width (`int`): Width of the output video frames.
+            height (`int`): Height of the output video frames.
+            seed (`int`): Random seed for reproducibility.
+
+        Returns:
+            `CogVideoXDDIMInversionOutput`: Contains the inverse latents and reconstructed latents.
+        """
+        if not self.transformer.config.use_rotary_positional_embeddings:
+            raise NotImplementedError("This script supports CogVideoX 5B model only.")
+        video_frames = get_video_frames(
+            video_path=video_path,
+            width=width,
+            height=height,
+            skip_frames_start=skip_frames_start,
+            skip_frames_end=skip_frames_end,
+            max_num_frames=max_num_frames,
+            frame_sample_step=frame_sample_step,
+        ).to(device=self.device)
+        video_latents = self.encode_video_frames(video_frames=video_frames)
+        inverse_latents = self.sample(
+            latents=video_latents,
+            scheduler=self.inverse_scheduler,
+            prompt="",
+            num_inference_steps=num_inference_steps,
+            guidance_scale=guidance_scale,
+            generator=torch.Generator(device=self.device).manual_seed(seed),
+        )
+        with OverrideAttnProcessors(transformer=self.transformer):
+            recon_latents = self.sample(
+                latents=torch.randn_like(video_latents),
+                scheduler=self.scheduler,
+                prompt=prompt,
+                num_inference_steps=num_inference_steps,
+                guidance_scale=guidance_scale,
+                generator=torch.Generator(device=self.device).manual_seed(seed),
+                reference_latents=reversed(inverse_latents),
+            )
+        return CogVideoXDDIMInversionOutput(
+            inverse_latents=inverse_latents,
+            recon_latents=recon_latents,
+        )
+
+
+if __name__ == "__main__":
+    arguments = get_args()
+    pipeline = CogVideoXPipelineForDDIMInversion.from_pretrained(
+        arguments.pop("model_path"),
+        torch_dtype=arguments.pop("dtype"),
+    ).to(device=arguments.pop("device"))
+
+    output_path = arguments.pop("output_path")
+    fps = arguments.pop("fps")
+    inverse_video_path = os.path.join(output_path, f"{arguments.get('video_path')}_inversion.mp4")
+    recon_video_path = os.path.join(output_path, f"{arguments.get('video_path')}_reconstruction.mp4")
+
+    # Run DDIM inversion
+    output = pipeline(**arguments)
+    pipeline.export_latents_to_video(output.inverse_latents[-1], inverse_video_path, fps)
+    pipeline.export_latents_to_video(output.recon_latents[-1], recon_video_path, fps)

examples/community/pipeline_stg_cogvideox.py

@@ -0,0 +1,876 @@
+# Copyright 2024 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import CogVideoXLoraLoaderMixin
+from diffusers.models import AutoencoderKLCogVideoX, CogVideoXTransformer3DModel
+from diffusers.models.embeddings import get_3d_rotary_pos_embed
+from diffusers.pipelines.cogvideo.pipeline_output import CogVideoXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import CogVideoXDDIMScheduler, CogVideoXDPMScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from examples.community.pipeline_stg_cogvideox import CogVideoXSTGPipeline
+
+        >>> # Models: "THUDM/CogVideoX-2b" or "THUDM/CogVideoX-5b"
+        >>> pipe = CogVideoXSTGPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.float16).to("cuda")
+        >>> prompt = (
+        ...     "A father and son building a treehouse together, their hands covered in sawdust and smiles on their faces, realistic style."
+        ... )
+        >>> pipe.transformer.to(memory_format=torch.channels_last)
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [11]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0  # Set to 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> # Generate video frames with STG parameters
+        >>> frames = pipe(
+        ...     prompt=prompt,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling,
+        >>> ).frames[0]
+        >>> export_to_video(frames, "output.mp4", fps=8)
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+) -> torch.Tensor:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+
+    text_seq_length = encoder_hidden_states.size(1)
+
+    # norm & modulate
+    norm_hidden_states, norm_encoder_hidden_states, gate_msa, enc_gate_msa = self.norm1(
+        hidden_states, encoder_hidden_states, temb
+    )
+
+    # attention
+    attn_hidden_states, attn_encoder_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        image_rotary_emb=image_rotary_emb,
+    )
+
+    hidden_states = hidden_states + gate_msa * attn_hidden_states
+    encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder_hidden_states
+
+    # norm & modulate
+    norm_hidden_states, norm_encoder_hidden_states, gate_ff, enc_gate_ff = self.norm2(
+        hidden_states, encoder_hidden_states, temb
+    )
+
+    # feed-forward
+    norm_hidden_states = torch.cat([norm_encoder_hidden_states, norm_hidden_states], dim=1)
+    ff_output = self.ff(norm_hidden_states)
+
+    hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+    encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+    hidden_states[2:] = hidden_states_ptb
+    encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states, encoder_hidden_states
+
+
+# Similar to diffusers.pipelines.hunyuandit.pipeline_hunyuandit.get_resize_crop_region_for_grid
+def get_resize_crop_region_for_grid(src, tgt_width, tgt_height):
+    tw = tgt_width
+    th = tgt_height
+    h, w = src
+    r = h / w
+    if r > (th / tw):
+        resize_height = th
+        resize_width = int(round(th / h * w))
+    else:
+        resize_width = tw
+        resize_height = int(round(tw / w * h))
+
+    crop_top = int(round((th - resize_height) / 2.0))
+    crop_left = int(round((tw - resize_width) / 2.0))
+
+    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class CogVideoXSTGPipeline(DiffusionPipeline, CogVideoXLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using CogVideoX.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            Frozen text-encoder. CogVideoX uses
+            [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel); specifically the
+            [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+        tokenizer (`T5Tokenizer`):
+            Tokenizer of class
+            [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+        transformer ([`CogVideoXTransformer3DModel`]):
+            A text conditioned `CogVideoXTransformer3DModel` to denoise the encoded video latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded video latents.
+    """
+
+    _optional_components = []
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+    _callback_tensor_inputs = [
+        "latents",
+        "prompt_embeds",
+        "negative_prompt_embeds",
+    ]
+
+    def __init__(
+        self,
+        tokenizer: T5Tokenizer,
+        text_encoder: T5EncoderModel,
+        vae: AutoencoderKLCogVideoX,
+        transformer: CogVideoXTransformer3DModel,
+        scheduler: Union[CogVideoXDDIMScheduler, CogVideoXDPMScheduler],
+    ):
+        super().__init__()
+
+        self.register_modules(
+            tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+        )
+        self.vae_scale_factor_spatial = (
+            2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
+        )
+        self.vae_scale_factor_temporal = (
+            self.vae.config.temporal_compression_ratio if getattr(self, "vae", None) else 4
+        )
+        self.vae_scaling_factor_image = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.7
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 226,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def prepare_latents(
+        self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+    ):
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        shape = (
+            batch_size,
+            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
+            num_channels_latents,
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+        )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def decode_latents(self, latents: torch.Tensor) -> torch.Tensor:
+        latents = latents.permute(0, 2, 1, 3, 4)  # [batch_size, num_channels, num_frames, height, width]
+        latents = 1 / self.vae_scaling_factor_image * latents
+
+        frames = self.vae.decode(latents).sample
+        return frames
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    # Copied from diffusers.pipelines.latte.pipeline_latte.LattePipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        negative_prompt,
+        callback_on_step_end_tensor_inputs,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+    def fuse_qkv_projections(self) -> None:
+        r"""Enables fused QKV projections."""
+        self.fusing_transformer = True
+        self.transformer.fuse_qkv_projections()
+
+    def unfuse_qkv_projections(self) -> None:
+        r"""Disable QKV projection fusion if enabled."""
+        if not self.fusing_transformer:
+            logger.warning("The Transformer was not initially fused for QKV projections. Doing nothing.")
+        else:
+            self.transformer.unfuse_qkv_projections()
+            self.fusing_transformer = False
+
+    def _prepare_rotary_positional_embeddings(
+        self,
+        height: int,
+        width: int,
+        num_frames: int,
+        device: torch.device,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        grid_height = height // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+        grid_width = width // (self.vae_scale_factor_spatial * self.transformer.config.patch_size)
+
+        p = self.transformer.config.patch_size
+        p_t = self.transformer.config.patch_size_t
+
+        base_size_width = self.transformer.config.sample_width // p
+        base_size_height = self.transformer.config.sample_height // p
+
+        if p_t is None:
+            # CogVideoX 1.0
+            grid_crops_coords = get_resize_crop_region_for_grid(
+                (grid_height, grid_width), base_size_width, base_size_height
+            )
+            freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+                embed_dim=self.transformer.config.attention_head_dim,
+                crops_coords=grid_crops_coords,
+                grid_size=(grid_height, grid_width),
+                temporal_size=num_frames,
+                device=device,
+            )
+        else:
+            # CogVideoX 1.5
+            base_num_frames = (num_frames + p_t - 1) // p_t
+
+            freqs_cos, freqs_sin = get_3d_rotary_pos_embed(
+                embed_dim=self.transformer.config.attention_head_dim,
+                crops_coords=None,
+                grid_size=(grid_height, grid_width),
+                temporal_size=base_num_frames,
+                grid_type="slice",
+                max_size=(base_size_height, base_size_width),
+                device=device,
+            )
+
+        return freqs_cos, freqs_sin
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_frames: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: Optional[List[int]] = None,
+        guidance_scale: float = 6,
+        use_dynamic_cfg: bool = False,
+        num_videos_per_prompt: int = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: str = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 226,
+        stg_applied_layers_idx: Optional[List[int]] = [11],
+        stg_scale: Optional[float] = 0.0,
+        do_rescaling: Optional[bool] = False,
+    ) -> Union[CogVideoXPipelineOutput, Tuple]:
+        """
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            height (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, *optional*, defaults to self.transformer.config.sample_height * self.vae_scale_factor_spatial):
+                The width in pixels of the generated image. This is set to 720 by default for the best results.
+            num_frames (`int`, defaults to `48`):
+                Number of frames to generate. Must be divisible by self.vae_scale_factor_temporal. Generated video will
+                contain 1 extra frame because CogVideoX is conditioned with (num_seconds * fps + 1) frames where
+                num_seconds is 6 and fps is 8. However, since videos can be saved at any fps, the only condition that
+                needs to be satisfied is that of divisibility mentioned above.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+                of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int`, defaults to `226`):
+                Maximum sequence length in encoded prompt. Must be consistent with
+                `self.transformer.config.max_text_seq_length` otherwise may lead to poor results.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] or `tuple`:
+            [`~pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipelineOutput`] if `return_dict` is True, otherwise a
+            `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.transformer.config.sample_height * self.vae_scale_factor_spatial
+        width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial
+        num_frames = num_frames or self.transformer.config.sample_frames
+
+        num_videos_per_prompt = 1
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            negative_prompt,
+            callback_on_step_end_tensor_inputs,
+            prompt_embeds,
+            negative_prompt_embeds,
+        )
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Default call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale > 1.0
+
+        # 3. Encode input prompt
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            negative_prompt,
+            do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+        elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+        self._num_timesteps = len(timesteps)
+
+        # 5. Prepare latents
+        latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+
+        # For CogVideoX 1.5, the latent frames should be padded to make it divisible by patch_size_t
+        patch_size_t = self.transformer.config.patch_size_t
+        additional_frames = 0
+        if patch_size_t is not None and latent_frames % patch_size_t != 0:
+            additional_frames = patch_size_t - latent_frames % patch_size_t
+            num_frames += additional_frames * self.vae_scale_factor_temporal
+
+        latent_channels = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            latent_channels,
+            num_frames,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7. Create rotary embeds if required
+        image_rotary_emb = (
+            self._prepare_rotary_positional_embeddings(height, width, latents.size(1), device)
+            if self.transformer.config.use_rotary_positional_embeddings
+            else None
+        )
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            # for DPM-solver++
+            old_pred_original_sample = None
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                # predict noise model_output
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    image_rotary_emb=image_rotary_emb,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                # perform guidance
+                if use_dynamic_cfg:
+                    self._guidance_scale = 1 + guidance_scale * (
+                        (1 - math.cos(math.pi * ((num_inference_steps - t.item()) / num_inference_steps) ** 5.0)) / 2
+                    )
+                if do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                elif do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                if not isinstance(self.scheduler, CogVideoXDPMScheduler):
+                    latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+                else:
+                    latents, old_pred_original_sample = self.scheduler.step(
+                        noise_pred,
+                        old_pred_original_sample,
+                        t,
+                        timesteps[i - 1] if i > 0 else None,
+                        latents,
+                        **extra_step_kwargs,
+                        return_dict=False,
+                    )
+                latents = latents.to(prompt_embeds.dtype)
+
+                # call the callback, if provided
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            # Discard any padding frames that were added for CogVideoX 1.5
+            latents = latents[:, additional_frames:]
+            video = self.decode_latents(latents)
+            video = self.video_processor.postprocess_video(video=video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return CogVideoXPipelineOutput(frames=video)

examples/community/pipeline_stg_hunyuan_video.py

@@ -0,0 +1,794 @@
+# Copyright 2024 The HunyuanVideo Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer, LlamaModel, LlamaTokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import HunyuanVideoLoraLoaderMixin
+from diffusers.models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel
+from diffusers.pipelines.hunyuan_video.pipeline_output import HunyuanVideoPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from diffusers import HunyuanVideoTransformer3DModel
+        >>> from examples.community.pipeline_stg_hunyuan_video import HunyuanVideoSTGPipeline
+
+        >>> model_id = "hunyuanvideo-community/HunyuanVideo"
+        >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+        ...     model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe = HunyuanVideoSTGPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.float16)
+        >>> pipe.vae.enable_tiling()
+        >>> pipe.to("cuda")
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [2]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+
+        >>> output = pipe(
+        ...     prompt="A wolf howling at the moon, with the moon subtly resembling a giant clock face, realistic style.",
+        ...     height=320,
+        ...     width=512,
+        ...     num_frames=61,
+        ...     num_inference_steps=30,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        >>> ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=15)
+        ```
+"""
+
+
+DEFAULT_PROMPT_TEMPLATE = {
+    "template": (
+        "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: "
+        "1. The main content and theme of the video."
+        "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects."
+        "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects."
+        "4. background environment, light, style and atmosphere."
+        "5. camera angles, movements, and transitions used in the video:<|eot_id|>"
+        "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+    ),
+    "crop_start": 95,
+}
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    return hidden_states, encoder_hidden_states
+
+
+def forward_without_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    # 1. Input normalization
+    norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+    norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+        encoder_hidden_states, emb=temb
+    )
+
+    # 2. Joint attention
+    attn_output, context_attn_output = self.attn(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        attention_mask=attention_mask,
+        image_rotary_emb=freqs_cis,
+    )
+
+    # 3. Modulation and residual connection
+    hidden_states = hidden_states + attn_output * gate_msa.unsqueeze(1)
+    encoder_hidden_states = encoder_hidden_states + context_attn_output * c_gate_msa.unsqueeze(1)
+
+    norm_hidden_states = self.norm2(hidden_states)
+    norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+
+    norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+    norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+    # 4. Feed-forward
+    ff_output = self.ff(norm_hidden_states)
+    context_ff_output = self.ff_context(norm_encoder_hidden_states)
+
+    hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output
+    encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+    return hidden_states, encoder_hidden_states
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class HunyuanVideoSTGPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation using HunyuanVideo.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        text_encoder ([`LlamaModel`]):
+            [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
+        tokenizer (`LlamaTokenizer`):
+            Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
+        transformer ([`HunyuanVideoTransformer3DModel`]):
+            Conditional Transformer to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLHunyuanVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder_2 ([`CLIPTextModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer_2 (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        text_encoder: LlamaModel,
+        tokenizer: LlamaTokenizerFast,
+        transformer: HunyuanVideoTransformer3DModel,
+        vae: AutoencoderKLHunyuanVideo,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        text_encoder_2: CLIPTextModel,
+        tokenizer_2: CLIPTokenizer,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+            text_encoder_2=text_encoder_2,
+            tokenizer_2=tokenizer_2,
+        )
+
+        self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_llama_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_template: Dict[str, Any],
+        num_videos_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+        num_hidden_layers_to_skip: int = 2,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        prompt = [prompt_template["template"].format(p) for p in prompt]
+
+        crop_start = prompt_template.get("crop_start", None)
+        if crop_start is None:
+            prompt_template_input = self.tokenizer(
+                prompt_template["template"],
+                padding="max_length",
+                return_tensors="pt",
+                return_length=False,
+                return_overflowing_tokens=False,
+                return_attention_mask=False,
+            )
+            crop_start = prompt_template_input["input_ids"].shape[-1]
+            # Remove <|eot_id|> token and placeholder {}
+            crop_start -= 2
+
+        max_sequence_length += crop_start
+        text_inputs = self.tokenizer(
+            prompt,
+            max_length=max_sequence_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_attention_mask=True,
+        )
+        text_input_ids = text_inputs.input_ids.to(device=device)
+        prompt_attention_mask = text_inputs.attention_mask.to(device=device)
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_attention_mask,
+            output_hidden_states=True,
+        ).hidden_states[-(num_hidden_layers_to_skip + 1)]
+        prompt_embeds = prompt_embeds.to(dtype=dtype)
+
+        if crop_start is not None and crop_start > 0:
+            prompt_embeds = prompt_embeds[:, crop_start:]
+            prompt_attention_mask = prompt_attention_mask[:, crop_start:]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt)
+        prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_attention_mask
+
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_videos_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 77,
+    ) -> torch.Tensor:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder_2.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer_2(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because CLIP can only handle sequences up to"
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        prompt_2: Union[str, List[str]] = None,
+        prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+    ):
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
+                prompt,
+                prompt_template,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=max_sequence_length,
+            )
+
+        if pooled_prompt_embeds is None:
+            if prompt_2 is None and pooled_prompt_embeds is None:
+                prompt_2 = prompt
+            pooled_prompt_embeds = self._get_clip_prompt_embeds(
+                prompt,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=77,
+            )
+
+        return prompt_embeds, pooled_prompt_embeds, prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_template=None,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if prompt_template is not None:
+            if not isinstance(prompt_template, dict):
+                raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}")
+            if "template" not in prompt_template:
+                raise ValueError(
+                    f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}"
+                )
+
+    def prepare_latents(
+        self,
+        batch_size: int,
+        num_channels_latents: 32,
+        height: int = 720,
+        width: int = 1280,
+        num_frames: int = 129,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        shape = (
+            batch_size,
+            num_channels_latents,
+            num_frames,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Union[str, List[str]] = None,
+        height: int = 720,
+        width: int = 1280,
+        num_frames: int = 129,
+        num_inference_steps: int = 50,
+        sigmas: List[float] = None,
+        guidance_scale: float = 6.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
+        max_sequence_length: int = 256,
+        stg_applied_layers_idx: Optional[List[int]] = [2],
+        stg_scale: Optional[float] = 0.0,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead.
+            height (`int`, defaults to `720`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `1280`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `129`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            guidance_scale (`float`, defaults to `6.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality. Note that the only available HunyuanVideo model is
+                CFG-distilled, which means that traditional guidance between unconditional and conditional latent is
+                not applied.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~HunyuanVideoPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            prompt_template,
+        )
+
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Encode input prompt
+        prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_template=prompt_template,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            device=device,
+            max_sequence_length=max_sequence_length,
+        )
+
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        prompt_attention_mask = prompt_attention_mask.to(transformer_dtype)
+        if pooled_prompt_embeds is not None:
+            pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype)
+
+        # 4. Prepare timesteps
+        sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            sigmas=sigmas,
+        )
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_latent_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. Prepare guidance condition
+        guidance = torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                latent_model_input = latents.to(transformer_dtype)
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                if self.do_spatio_temporal_guidance:
+                    for i in stg_applied_layers_idx:
+                        self.transformer.transformer_blocks[i].forward = types.MethodType(
+                            forward_without_stg, self.transformer.transformer_blocks[i]
+                        )
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    pooled_projections=pooled_prompt_embeds,
+                    guidance=guidance,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_spatio_temporal_guidance:
+                    for i in stg_applied_layers_idx:
+                        self.transformer.transformer_blocks[i].forward = types.MethodType(
+                            forward_with_stg, self.transformer.transformer_blocks[i]
+                        )
+
+                    noise_pred_perturb = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=prompt_embeds,
+                        encoder_attention_mask=prompt_attention_mask,
+                        pooled_projections=pooled_prompt_embeds,
+                        guidance=guidance,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = noise_pred + self._stg_scale * (noise_pred - noise_pred_perturb)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype) / self.vae.config.scaling_factor
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+        else:
+            video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return HunyuanVideoPipelineOutput(frames=video)

examples/community/pipeline_stg_ltx.py

@@ -0,0 +1,886 @@
+# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKLLTXVideo
+from diffusers.models.transformers import LTXVideoTransformer3DModel
+from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from examples.community.pipeline_stg_ltx import LTXSTGPipeline
+
+        >>> pipe = LTXSTGPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A woman with light skin, wearing a blue jacket and a black hat with a veil, looks down and to her right, then back up as she speaks; she has brown hair styled in an updo, light brown eyebrows, and is wearing a white collared shirt under her jacket; the camera remains stationary on her face as she speaks; the background is out of focus, but shows trees and people in period clothing; the scene is captured in real-life footage."
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [19]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> video = pipe(
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=704,
+        ...     height=480,
+        ...     num_frames=161,
+        ...     num_inference_steps=50,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling,
+        >>> ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    encoder_attention_mask: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+
+    batch_size = hidden_states.size(0)
+    norm_hidden_states = self.norm1(hidden_states)
+
+    num_ada_params = self.scale_shift_table.shape[0]
+    ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
+    shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
+    norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+    attn_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=None,
+        image_rotary_emb=image_rotary_emb,
+    )
+    hidden_states = hidden_states + attn_hidden_states * gate_msa
+
+    attn_hidden_states = self.attn2(
+        hidden_states,
+        encoder_hidden_states=encoder_hidden_states,
+        image_rotary_emb=None,
+        attention_mask=encoder_attention_mask,
+    )
+    hidden_states = hidden_states + attn_hidden_states
+    norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
+
+    ff_output = self.ff(norm_hidden_states)
+    hidden_states = hidden_states + ff_output * gate_mlp
+
+    hidden_states[2:] = hidden_states_ptb
+    encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class LTXSTGPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
+    r"""
+    Pipeline for text-to-video generation.
+
+    Reference: https://github.com/Lightricks/LTX-Video
+
+    Args:
+        transformer ([`LTXVideoTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLLTXVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_spatial_compression_ratio = (
+            self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
+        )
+        self.vae_temporal_compression_ratio = (
+            self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
+        )
+        self.transformer_spatial_patch_size = (
+            self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
+        )
+        self.transformer_temporal_patch_size = (
+            self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128
+        )
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    @staticmethod
+    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
+        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
+        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
+        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
+        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
+        batch_size, num_channels, num_frames, height, width = latents.shape
+        post_patch_num_frames = num_frames // patch_size_t
+        post_patch_height = height // patch_size
+        post_patch_width = width // patch_size
+        latents = latents.reshape(
+            batch_size,
+            -1,
+            post_patch_num_frames,
+            patch_size_t,
+            post_patch_height,
+            patch_size,
+            post_patch_width,
+            patch_size,
+        )
+        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
+        return latents
+
+    @staticmethod
+    def _unpack_latents(
+        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
+    ) -> torch.Tensor:
+        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
+        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
+        # what happens in the `_pack_latents` method.
+        batch_size = latents.size(0)
+        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
+        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
+        return latents
+
+    @staticmethod
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def prepare_latents(
+        self,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+
+        height = height // self.vae_spatial_compression_ratio
+        width = width // self.vae_spatial_compression_ratio
+        num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        decode_timestep: Union[float, List[float]] = 0.0,
+        decode_noise_scale: Optional[Union[float, List[float]]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 128,
+        stg_applied_layers_idx: Optional[List[int]] = [19],
+        stg_scale: Optional[float] = 1.0,
+        do_rescaling: Optional[bool] = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `3 `):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            decode_timestep (`float`, defaults to `0.0`):
+                The timestep at which generated video is decoded.
+            decode_noise_scale (`float`, defaults to `None`):
+                The interpolation factor between random noise and denoised latents at the decode timestep.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat(
+                [negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
+            )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timesteps
+        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+        latent_height = height // self.vae_spatial_compression_ratio
+        latent_width = width // self.vae_spatial_compression_ratio
+        video_sequence_length = latent_num_frames * latent_height * latent_width
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        mu = calculate_shift(
+            video_sequence_length,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.16),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare micro-conditions
+        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
+        rope_interpolation_scale = (
+            1 / latent_frame_rate,
+            self.vae_spatial_compression_ratio,
+            self.vae_spatial_compression_ratio,
+        )
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    num_frames=latent_num_frames,
+                    height=latent_height,
+                    width=latent_width,
+                    rope_interpolation_scale=rope_interpolation_scale,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+        else:
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            latents = self._denormalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            latents = latents.to(prompt_embeds.dtype)
+
+            if not self.vae.config.timestep_conditioning:
+                timestep = None
+            else:
+                noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
+                if not isinstance(decode_timestep, list):
+                    decode_timestep = [decode_timestep] * batch_size
+                if decode_noise_scale is None:
+                    decode_noise_scale = decode_timestep
+                elif not isinstance(decode_noise_scale, list):
+                    decode_noise_scale = [decode_noise_scale] * batch_size
+
+                timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
+                decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
+                    :, None, None, None, None
+                ]
+                latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
+
+            video = self.vae.decode(latents, timestep, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)

examples/community/pipeline_stg_ltx_image2video.py

@@ -0,0 +1,985 @@
+# Copyright 2024 Lightricks and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.image_processor import PipelineImageInput
+from diffusers.loaders import FromSingleFileMixin, LTXVideoLoraLoaderMixin
+from diffusers.models.autoencoders import AutoencoderKLLTXVideo
+from diffusers.models.transformers import LTXVideoTransformer3DModel
+from diffusers.pipelines.ltx.pipeline_output import LTXPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import export_to_video, load_image
+        >>> from examples.community.pipeline_stg_ltx_image2video import LTXImageToVideoSTGPipeline
+
+        >>> pipe = LTXImageToVideoSTGPipeline.from_pretrained("Lightricks/LTX-Video", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/a-r-r-o-w/tiny-meme-dataset-captioned/resolve/main/images/11.png"
+        >>> )
+        >>> prompt = "A medieval fantasy scene featuring a rugged man with shoulder-length brown hair and a beard. He wears a dark leather tunic over a maroon shirt with intricate metal details. His facial expression is serious and intense, and he is making a gesture with his right hand, forming a small circle with his thumb and index finger. The warm golden lighting casts dramatic shadows on his face. The background includes an ornate stone arch and blurred medieval-style decor, creating an epic atmosphere."
+        >>> negative_prompt = "worst quality, inconsistent motion, blurry, jittery, distorted"
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [19]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> video = pipe(
+        ...     image=image,
+        ...     prompt=prompt,
+        ...     negative_prompt=negative_prompt,
+        ...     width=704,
+        ...     height=480,
+        ...     num_frames=161,
+        ...     num_inference_steps=50,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling,
+        >>> ).frames[0]
+        >>> export_to_video(video, "output.mp4", fps=24)
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    encoder_attention_mask: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+
+    batch_size = hidden_states.size(0)
+    norm_hidden_states = self.norm1(hidden_states)
+
+    num_ada_params = self.scale_shift_table.shape[0]
+    ada_values = self.scale_shift_table[None, None] + temb.reshape(batch_size, temb.size(1), num_ada_params, -1)
+    shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ada_values.unbind(dim=2)
+    norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+    attn_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=None,
+        image_rotary_emb=image_rotary_emb,
+    )
+    hidden_states = hidden_states + attn_hidden_states * gate_msa
+
+    attn_hidden_states = self.attn2(
+        hidden_states,
+        encoder_hidden_states=encoder_hidden_states,
+        image_rotary_emb=None,
+        attention_mask=encoder_attention_mask,
+    )
+    hidden_states = hidden_states + attn_hidden_states
+    norm_hidden_states = self.norm2(hidden_states) * (1 + scale_mlp) + shift_mlp
+
+    ff_output = self.ff(norm_hidden_states)
+    hidden_states = hidden_states + ff_output * gate_mlp
+
+    hidden_states[2:] = hidden_states_ptb
+    encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class LTXImageToVideoSTGPipeline(DiffusionPipeline, FromSingleFileMixin, LTXVideoLoraLoaderMixin):
+    r"""
+    Pipeline for image-to-video generation.
+
+    Reference: https://github.com/Lightricks/LTX-Video
+
+    Args:
+        transformer ([`LTXVideoTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLLTXVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLLTXVideo,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: LTXVideoTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+
+        self.vae_spatial_compression_ratio = (
+            self.vae.spatial_compression_ratio if getattr(self, "vae", None) is not None else 32
+        )
+        self.vae_temporal_compression_ratio = (
+            self.vae.temporal_compression_ratio if getattr(self, "vae", None) is not None else 8
+        )
+        self.transformer_spatial_patch_size = (
+            self.transformer.config.patch_size if getattr(self, "transformer", None) is not None else 1
+        )
+        self.transformer_temporal_patch_size = (
+            self.transformer.config.patch_size_t if getattr(self, "transformer") is not None else 1
+        )
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_compression_ratio)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if getattr(self, "tokenizer", None) is not None else 128
+        )
+
+        self.default_height = 512
+        self.default_width = 704
+        self.default_frames = 121
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device))[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.mochi.pipeline_mochi.MochiPipeline.encode_prompt with 256->128
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 128,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline.check_inputs
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 32 != 0 or width % 32 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._pack_latents
+    def _pack_latents(latents: torch.Tensor, patch_size: int = 1, patch_size_t: int = 1) -> torch.Tensor:
+        # Unpacked latents of shape are [B, C, F, H, W] are patched into tokens of shape [B, C, F // p_t, p_t, H // p, p, W // p, p].
+        # The patch dimensions are then permuted and collapsed into the channel dimension of shape:
+        # [B, F // p_t * H // p * W // p, C * p_t * p * p] (an ndim=3 tensor).
+        # dim=0 is the batch size, dim=1 is the effective video sequence length, dim=2 is the effective number of input features
+        batch_size, num_channels, num_frames, height, width = latents.shape
+        post_patch_num_frames = num_frames // patch_size_t
+        post_patch_height = height // patch_size
+        post_patch_width = width // patch_size
+        latents = latents.reshape(
+            batch_size,
+            -1,
+            post_patch_num_frames,
+            patch_size_t,
+            post_patch_height,
+            patch_size,
+            post_patch_width,
+            patch_size,
+        )
+        latents = latents.permute(0, 2, 4, 6, 1, 3, 5, 7).flatten(4, 7).flatten(1, 3)
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._unpack_latents
+    def _unpack_latents(
+        latents: torch.Tensor, num_frames: int, height: int, width: int, patch_size: int = 1, patch_size_t: int = 1
+    ) -> torch.Tensor:
+        # Packed latents of shape [B, S, D] (S is the effective video sequence length, D is the effective feature dimensions)
+        # are unpacked and reshaped into a video tensor of shape [B, C, F, H, W]. This is the inverse operation of
+        # what happens in the `_pack_latents` method.
+        batch_size = latents.size(0)
+        latents = latents.reshape(batch_size, num_frames, height, width, -1, patch_size_t, patch_size, patch_size)
+        latents = latents.permute(0, 4, 1, 5, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(2, 3)
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._normalize_latents
+    def _normalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Normalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = (latents - latents_mean) * scaling_factor / latents_std
+        return latents
+
+    @staticmethod
+    # Copied from diffusers.pipelines.ltx.pipeline_ltx.LTXPipeline._denormalize_latents
+    def _denormalize_latents(
+        latents: torch.Tensor, latents_mean: torch.Tensor, latents_std: torch.Tensor, scaling_factor: float = 1.0
+    ) -> torch.Tensor:
+        # Denormalize latents across the channel dimension [B, C, F, H, W]
+        latents_mean = latents_mean.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents_std = latents_std.view(1, -1, 1, 1, 1).to(latents.device, latents.dtype)
+        latents = latents * latents_std / scaling_factor + latents_mean
+        return latents
+
+    def prepare_latents(
+        self,
+        image: Optional[torch.Tensor] = None,
+        batch_size: int = 1,
+        num_channels_latents: int = 128,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        height = height // self.vae_spatial_compression_ratio
+        width = width // self.vae_spatial_compression_ratio
+        num_frames = (
+            (num_frames - 1) // self.vae_temporal_compression_ratio + 1 if latents is None else latents.size(2)
+        )
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+        mask_shape = (batch_size, 1, num_frames, height, width)
+
+        if latents is not None:
+            conditioning_mask = latents.new_zeros(shape)
+            conditioning_mask[:, :, 0] = 1.0
+            conditioning_mask = self._pack_latents(
+                conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+            )
+            return latents.to(device=device, dtype=dtype), conditioning_mask
+
+        if isinstance(generator, list):
+            if len(generator) != batch_size:
+                raise ValueError(
+                    f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                    f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+                )
+
+            init_latents = [
+                retrieve_latents(self.vae.encode(image[i].unsqueeze(0).unsqueeze(2)), generator[i])
+                for i in range(batch_size)
+            ]
+        else:
+            init_latents = [
+                retrieve_latents(self.vae.encode(img.unsqueeze(0).unsqueeze(2)), generator) for img in image
+            ]
+
+        init_latents = torch.cat(init_latents, dim=0).to(dtype)
+        init_latents = self._normalize_latents(init_latents, self.vae.latents_mean, self.vae.latents_std)
+        init_latents = init_latents.repeat(1, 1, num_frames, 1, 1)
+        conditioning_mask = torch.zeros(mask_shape, device=device, dtype=dtype)
+        conditioning_mask[:, :, 0] = 1.0
+
+        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        latents = init_latents * conditioning_mask + noise * (1 - conditioning_mask)
+
+        conditioning_mask = self._pack_latents(
+            conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        ).squeeze(-1)
+        latents = self._pack_latents(
+            latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+        )
+
+        return latents, conditioning_mask
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PipelineImageInput = None,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: int = 512,
+        width: int = 704,
+        num_frames: int = 161,
+        frame_rate: int = 25,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        decode_timestep: Union[float, List[float]] = 0.0,
+        decode_noise_scale: Optional[Union[float, List[float]]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 128,
+        stg_applied_layers_idx: Optional[List[int]] = [19],
+        stg_scale: Optional[float] = 1.0,
+        do_rescaling: Optional[bool] = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            image (`PipelineImageInput`):
+                The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`.
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, defaults to `512`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, defaults to `704`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `161`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `3 `):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            decode_timestep (`float`, defaults to `0.0`):
+                The timestep at which generated video is decoded.
+            decode_noise_scale (`float`, defaults to `None`):
+                The interpolation factor between random noise and denoised latents at the decode timestep.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.ltx.LTXPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `128 `):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ltx.LTXPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.ltx.LTXPipelineOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._stg_scale = stg_scale
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat(
+                [negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
+            )
+
+        # 4. Prepare latent variables
+        if latents is None:
+            image = self.video_processor.preprocess(image, height=height, width=width)
+            image = image.to(device=device, dtype=prompt_embeds.dtype)
+
+        num_channels_latents = self.transformer.config.in_channels
+        latents, conditioning_mask = self.prepare_latents(
+            image,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            torch.float32,
+            device,
+            generator,
+            latents,
+        )
+
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            conditioning_mask = torch.cat([conditioning_mask, conditioning_mask])
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            conditioning_mask = torch.cat([conditioning_mask, conditioning_mask, conditioning_mask])
+
+        # 5. Prepare timesteps
+        latent_num_frames = (num_frames - 1) // self.vae_temporal_compression_ratio + 1
+        latent_height = height // self.vae_spatial_compression_ratio
+        latent_width = width // self.vae_spatial_compression_ratio
+        video_sequence_length = latent_num_frames * latent_height * latent_width
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        mu = calculate_shift(
+            video_sequence_length,
+            self.scheduler.config.get("base_image_seq_len", 256),
+            self.scheduler.config.get("max_image_seq_len", 4096),
+            self.scheduler.config.get("base_shift", 0.5),
+            self.scheduler.config.get("max_shift", 1.16),
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas=sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Prepare micro-conditions
+        latent_frame_rate = frame_rate / self.vae_temporal_compression_ratio
+        rope_interpolation_scale = (
+            1 / latent_frame_rate,
+            self.vae_spatial_compression_ratio,
+            self.vae_spatial_compression_ratio,
+        )
+
+        # 7. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+
+                latent_model_input = latent_model_input.to(prompt_embeds.dtype)
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0])
+                timestep = timestep.unsqueeze(-1) * (1 - conditioning_mask)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    num_frames=latent_num_frames,
+                    height=latent_height,
+                    width=latent_width,
+                    rope_interpolation_scale=rope_interpolation_scale,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                noise_pred = noise_pred.float()
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                    timestep, _ = timestep.chunk(2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+                    timestep, _, _ = timestep.chunk(3)
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                noise_pred = self._unpack_latents(
+                    noise_pred,
+                    latent_num_frames,
+                    latent_height,
+                    latent_width,
+                    self.transformer_spatial_patch_size,
+                    self.transformer_temporal_patch_size,
+                )
+                latents = self._unpack_latents(
+                    latents,
+                    latent_num_frames,
+                    latent_height,
+                    latent_width,
+                    self.transformer_spatial_patch_size,
+                    self.transformer_temporal_patch_size,
+                )
+
+                noise_pred = noise_pred[:, :, 1:]
+                noise_latents = latents[:, :, 1:]
+                pred_latents = self.scheduler.step(noise_pred, t, noise_latents, return_dict=False)[0]
+
+                latents = torch.cat([latents[:, :, :1], pred_latents], dim=2)
+                latents = self._pack_latents(
+                    latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
+                )
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+        else:
+            latents = self._unpack_latents(
+                latents,
+                latent_num_frames,
+                latent_height,
+                latent_width,
+                self.transformer_spatial_patch_size,
+                self.transformer_temporal_patch_size,
+            )
+            latents = self._denormalize_latents(
+                latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
+            )
+            latents = latents.to(prompt_embeds.dtype)
+
+            if not self.vae.config.timestep_conditioning:
+                timestep = None
+            else:
+                noise = torch.randn(latents.shape, generator=generator, device=device, dtype=latents.dtype)
+                if not isinstance(decode_timestep, list):
+                    decode_timestep = [decode_timestep] * batch_size
+                if decode_noise_scale is None:
+                    decode_noise_scale = decode_timestep
+                elif not isinstance(decode_noise_scale, list):
+                    decode_noise_scale = [decode_noise_scale] * batch_size
+
+                timestep = torch.tensor(decode_timestep, device=device, dtype=latents.dtype)
+                decode_noise_scale = torch.tensor(decode_noise_scale, device=device, dtype=latents.dtype)[
+                    :, None, None, None, None
+                ]
+                latents = (1 - decode_noise_scale) * latents + decode_noise_scale * noise
+
+            video = self.vae.decode(latents, timestep, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return LTXPipelineOutput(frames=video)

examples/community/pipeline_stg_mochi.py

@@ -0,0 +1,843 @@
+# Copyright 2024 Genmo and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import types
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import Mochi1LoraLoaderMixin
+from diffusers.models import AutoencoderKLMochi, MochiTransformer3DModel
+from diffusers.pipelines.mochi.pipeline_output import MochiPipelineOutput
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import (
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers.utils import export_to_video
+        >>> from examples.community.pipeline_stg_mochi import MochiSTGPipeline
+
+        >>> pipe = MochiSTGPipeline.from_pretrained("genmo/mochi-1-preview", torch_dtype=torch.bfloat16)
+        >>> pipe.enable_model_cpu_offload()
+        >>> pipe.enable_vae_tiling()
+        >>> prompt = "A close-up of a beautiful woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues, realistic style."
+
+        >>> # Configure STG mode options
+        >>> stg_applied_layers_idx = [34]  # Layer indices from 0 to 41
+        >>> stg_scale = 1.0 # Set 0.0 for CFG
+        >>> do_rescaling = False
+
+        >>> frames = pipe(
+        ...     prompt=prompt,
+        ...     num_inference_steps=28,
+        ...     guidance_scale=3.5,
+        ...     stg_applied_layers_idx=stg_applied_layers_idx,
+        ...     stg_scale=stg_scale,
+        ...     do_rescaling=do_rescaling).frames[0]
+        >>> export_to_video(frames, "mochi.mp4")
+        ```
+"""
+
+
+def forward_with_stg(
+    self,
+    hidden_states: torch.Tensor,
+    encoder_hidden_states: torch.Tensor,
+    temb: torch.Tensor,
+    encoder_attention_mask: torch.Tensor,
+    image_rotary_emb: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    hidden_states_ptb = hidden_states[2:]
+    encoder_hidden_states_ptb = encoder_hidden_states[2:]
+    norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+
+    if not self.context_pre_only:
+        norm_encoder_hidden_states, enc_gate_msa, enc_scale_mlp, enc_gate_mlp = self.norm1_context(
+            encoder_hidden_states, temb
+        )
+    else:
+        norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+
+    attn_hidden_states, context_attn_hidden_states = self.attn1(
+        hidden_states=norm_hidden_states,
+        encoder_hidden_states=norm_encoder_hidden_states,
+        image_rotary_emb=image_rotary_emb,
+        attention_mask=encoder_attention_mask,
+    )
+
+    hidden_states = hidden_states + self.norm2(attn_hidden_states, torch.tanh(gate_msa).unsqueeze(1))
+    norm_hidden_states = self.norm3(hidden_states, (1 + scale_mlp.unsqueeze(1).to(torch.float32)))
+    ff_output = self.ff(norm_hidden_states)
+    hidden_states = hidden_states + self.norm4(ff_output, torch.tanh(gate_mlp).unsqueeze(1))
+
+    if not self.context_pre_only:
+        encoder_hidden_states = encoder_hidden_states + self.norm2_context(
+            context_attn_hidden_states, torch.tanh(enc_gate_msa).unsqueeze(1)
+        )
+        norm_encoder_hidden_states = self.norm3_context(
+            encoder_hidden_states, (1 + enc_scale_mlp.unsqueeze(1).to(torch.float32))
+        )
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states + self.norm4_context(
+            context_ff_output, torch.tanh(enc_gate_mlp).unsqueeze(1)
+        )
+
+        hidden_states[2:] = hidden_states_ptb
+        encoder_hidden_states[2:] = encoder_hidden_states_ptb
+
+    return hidden_states, encoder_hidden_states
+
+
+# from: https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
+def linear_quadratic_schedule(num_steps, threshold_noise, linear_steps=None):
+    if linear_steps is None:
+        linear_steps = num_steps // 2
+    linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)]
+    threshold_noise_step_diff = linear_steps - threshold_noise * num_steps
+    quadratic_steps = num_steps - linear_steps
+    quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2)
+    linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2)
+    const = quadratic_coef * (linear_steps**2)
+    quadratic_sigma_schedule = [
+        quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps)
+    ]
+    sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule
+    sigma_schedule = [1.0 - x for x in sigma_schedule]
+    return sigma_schedule
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom value")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class MochiSTGPipeline(DiffusionPipeline, Mochi1LoraLoaderMixin):
+    r"""
+    The mochi pipeline for text-to-video generation.
+
+    Reference: https://github.com/genmoai/models
+
+    Args:
+        transformer ([`MochiTransformer3DModel`]):
+            Conditional Transformer architecture to denoise the encoded video latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLMochi`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKLMochi,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: MochiTransformer3DModel,
+        force_zeros_for_empty_prompt: bool = False,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        # TODO: determine these scaling factors from model parameters
+        self.vae_spatial_scale_factor = 8
+        self.vae_temporal_scale_factor = 6
+        self.patch_size = 2
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_spatial_scale_factor)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 256
+        )
+        self.default_height = 480
+        self.default_width = 848
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        prompt_attention_mask = prompt_attention_mask.bool().to(device)
+
+        # The original Mochi implementation zeros out empty negative prompts
+        # but this can lead to overflow when placing the entire pipeline under the autocast context
+        # adding this here so that we can enable zeroing prompts if necessary
+        if self.config.force_zeros_for_empty_prompt and (prompt == "" or prompt[-1] == ""):
+            text_input_ids = torch.zeros_like(text_input_ids, device=device)
+            prompt_attention_mask = torch.zeros_like(prompt_attention_mask, dtype=torch.bool, device=device)
+
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)[0]
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        prompt_attention_mask = prompt_attention_mask.view(batch_size, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Adapted from diffusers.pipelines.cogvideo.pipeline_cogvideox.CogVideoXPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        do_classifier_free_guidance: bool = True,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        max_sequence_length: int = 256,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+                Whether to use classifier free guidance or not.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                Number of videos that should be generated per prompt. torch device to place the resulting embeddings on
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            device: (`torch.device`, *optional*):
+                torch device
+            dtype: (`torch.dtype`, *optional*):
+                torch dtype
+        """
+        device = device or self._execution_device
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds, negative_prompt_attention_mask = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+                dtype=dtype,
+            )
+
+        return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        prompt_attention_mask=None,
+        negative_prompt_attention_mask=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+        if prompt_embeds is not None and prompt_attention_mask is None:
+            raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+        if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+            raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+            if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+                raise ValueError(
+                    "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+                    f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+                    f" {negative_prompt_attention_mask.shape}."
+                )
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        num_frames,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = height // self.vae_spatial_scale_factor
+        width = width // self.vae_spatial_scale_factor
+        num_frames = (num_frames - 1) // self.vae_temporal_scale_factor + 1
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=torch.float32)
+        latents = latents.to(dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def do_spatio_temporal_guidance(self):
+        return self._stg_scale > 0.0
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_frames: int = 19,
+        num_inference_steps: int = 64,
+        timesteps: List[int] = None,
+        guidance_scale: float = 4.5,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 256,
+        stg_applied_layers_idx: Optional[List[int]] = [34],
+        stg_scale: Optional[float] = 0.0,
+        do_rescaling: Optional[bool] = False,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, *optional*, defaults to `self.default_height`):
+                The height in pixels of the generated image. This is set to 480 by default for the best results.
+            width (`int`, *optional*, defaults to `self.default_width`):
+                The width in pixels of the generated image. This is set to 848 by default for the best results.
+            num_frames (`int`, defaults to `19`):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, defaults to `4.5`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of videos to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            prompt_attention_mask (`torch.Tensor`, *optional*):
+                Pre-generated attention mask for text embeddings.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not
+                provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+            negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+                Pre-generated attention mask for negative text embeddings.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.mochi.MochiPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to `256`):
+                Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.mochi.MochiPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.mochi.MochiPipelineOutput`] is returned, otherwise a `tuple`
+                is returned where the first element is a list with the generated images.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        height = height or self.default_height
+        width = width or self.default_width
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            height=height,
+            width=width,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._stg_scale = stg_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        if self.do_spatio_temporal_guidance:
+            for i in stg_applied_layers_idx:
+                self.transformer.transformer_blocks[i].forward = types.MethodType(
+                    forward_with_stg, self.transformer.transformer_blocks[i]
+                )
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+        # 3. Prepare text embeddings
+        (
+            prompt_embeds,
+            prompt_attention_mask,
+            negative_prompt_embeds,
+            negative_prompt_attention_mask,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            negative_prompt=negative_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            negative_prompt_attention_mask=negative_prompt_attention_mask,
+            max_sequence_length=max_sequence_length,
+            device=device,
+        )
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+        elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds], dim=0)
+            prompt_attention_mask = torch.cat(
+                [negative_prompt_attention_mask, prompt_attention_mask, prompt_attention_mask], dim=0
+            )
+
+        # 5. Prepare timestep
+        # from https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
+        threshold_noise = 0.025
+        sigmas = linear_quadratic_schedule(num_inference_steps, threshold_noise)
+        sigmas = np.array(sigmas)
+
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # Note: Mochi uses reversed timesteps. To ensure compatibility with methods like FasterCache, we need
+                # to make sure we're using the correct non-reversed timestep value.
+                self._current_timestep = 1000 - t
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 2)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    latent_model_input = torch.cat([latents] * 3)
+                else:
+                    latent_model_input = latents
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep=timestep,
+                    encoder_attention_mask=prompt_attention_mask,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                )[0]
+                # Mochi CFG + Sampling runs in FP32
+                noise_pred = noise_pred.to(torch.float32)
+
+                if self.do_classifier_free_guidance and not self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                elif self.do_classifier_free_guidance and self.do_spatio_temporal_guidance:
+                    noise_pred_uncond, noise_pred_text, noise_pred_perturb = noise_pred.chunk(3)
+                    noise_pred = (
+                        noise_pred_uncond
+                        + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+                        + self._stg_scale * (noise_pred_text - noise_pred_perturb)
+                    )
+
+                if do_rescaling:
+                    rescaling_scale = 0.7
+                    factor = noise_pred_text.std() / noise_pred.std()
+                    factor = rescaling_scale * factor + (1 - rescaling_scale)
+                    noise_pred = noise_pred * factor
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents.to(torch.float32), return_dict=False)[0]
+                latents = latents.to(latents_dtype)
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if output_type == "latent":
+            video = latents
+        else:
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 12, 1, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return MochiPipelineOutput(frames=video)

examples/custom_diffusion/train_custom_diffusion.py

@@ -1334,7 +1334,9 @@ def main(args):
 
         # run inference
         if args.validation_prompt and args.num_validation_images > 0:
-            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+            generator = (
+                torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
+            )
             images = [
                 pipeline(args.validation_prompt, num_inference_steps=25, generator=generator, eta=1.0).images[0]
                 for _ in range(args.num_validation_images)

examples/dreambooth/train_dreambooth_flux.py

@@ -172,7 +172,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     # autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
     autocast_ctx = nullcontext()
 

examples/dreambooth/train_dreambooth_lora.py

@@ -150,7 +150,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
 
     if args.validation_images is None:
         images = []
@@ -1119,17 +1119,22 @@ def main(args):
     )
 
     # Scheduler and math around the number of training steps.
-    overrode_max_train_steps = False
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1146,8 +1151,15 @@ def main(args):
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
+
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 

examples/dreambooth/train_dreambooth_lora_flux.py

@@ -181,7 +181,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     # autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
     autocast_ctx = nullcontext()
 

examples/dreambooth/train_dreambooth_lora_lumina2.py

@@ -167,7 +167,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
 
     with autocast_ctx:

examples/dreambooth/train_dreambooth_lora_sana.py

@@ -170,7 +170,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
 
     images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
 

examples/dreambooth/train_dreambooth_lora_sd3.py

@@ -199,7 +199,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     # autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
     autocast_ctx = nullcontext()
 

examples/dreambooth/train_dreambooth_lora_sdxl.py

@@ -203,17 +203,17 @@ def log_validation(
 
         pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args)
 
-    pipeline = pipeline.to(accelerator.device, dtype=torch_dtype)
+    pipeline = pipeline.to(accelerator.device)
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     # Currently the context determination is a bit hand-wavy. We can improve it in the future if there's a better
     # way to condition it. Reference: https://github.com/huggingface/diffusers/pull/7126#issuecomment-1968523051
     if torch.backends.mps.is_available() or "playground" in args.pretrained_model_name_or_path:
         autocast_ctx = nullcontext()
     else:
-        autocast_ctx = torch.autocast(accelerator.device.type)
+        autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
 
     with autocast_ctx:
         images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]

examples/dreambooth/train_dreambooth_sd3.py

@@ -175,7 +175,7 @@ def log_validation(
     pipeline.set_progress_bar_config(disable=True)
 
     # run inference
-    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed is not None else None
     # autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
     autocast_ctx = nullcontext()
 

examples/research_projects/anytext/README.md

@@ -0,0 +1,32 @@
+# AnyTextPipeline Pipeline
+
+Project page: https://aigcdesigngroup.github.io/homepage_anytext
+
+"AnyText comprises a diffusion pipeline with two primary elements: an auxiliary latent module and a text embedding module. The former uses inputs like text glyph, position, and masked image to generate latent features for text generation or editing. The latter employs an OCR model for encoding stroke data as embeddings, which blend with image caption embeddings from the tokenizer to generate texts that seamlessly integrate with the background. We employed text-control diffusion loss and text perceptual loss for training to further enhance writing accuracy."
+
+Each text line that needs to be generated should be enclosed in double quotes. For any usage questions, please refer to the [paper](https://arxiv.org/abs/2311.03054).
+
+
+```py
+import torch
+from diffusers import DiffusionPipeline
+from anytext_controlnet import AnyTextControlNetModel
+from diffusers.utils import load_image
+
+# I chose a font file shared by an HF staff:
+# !wget https://huggingface.co/spaces/ysharma/TranslateQuotesInImageForwards/resolve/main/arial-unicode-ms.ttf
+
+anytext_controlnet = AnyTextControlNetModel.from_pretrained("tolgacangoz/anytext-controlnet", torch_dtype=torch.float16,
+                                                            variant="fp16",)
+pipe = DiffusionPipeline.from_pretrained("tolgacangoz/anytext", font_path="arial-unicode-ms.ttf",
+                                          controlnet=anytext_controlnet, torch_dtype=torch.float16,
+                                          trust_remote_code=False,  # One needs to give permission to run this pipeline's code
+                                          ).to("cuda")
+
+# generate image
+prompt = 'photo of caramel macchiato coffee on the table, top-down perspective, with "Any" "Text" written on it using cream'
+draw_pos = load_image("https://raw.githubusercontent.com/tyxsspa/AnyText/refs/heads/main/example_images/gen9.png")
+image = pipe(prompt, num_inference_steps=20, mode="generate", draw_pos=draw_pos,
+             ).images[0]
+image
+```

examples/research_projects/anytext/anytext_controlnet.py

@@ -0,0 +1,463 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# Based on [AnyText: Multilingual Visual Text Generation And Editing](https://huggingface.co/papers/2311.03054).
+# Authors: Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, Xuansong Xie
+# Code: https://github.com/tyxsspa/AnyText with Apache-2.0 license
+#
+# Adapted to Diffusers by [M. Tolga Cangöz](https://github.com/tolgacangoz).
+
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from diffusers.configuration_utils import register_to_config
+from diffusers.models.controlnets.controlnet import (
+    ControlNetModel,
+    ControlNetOutput,
+)
+from diffusers.utils import logging
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class AnyTextControlNetConditioningEmbedding(nn.Module):
+    """
+    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+    model) to encode image-space conditions ... into feature maps ..."
+    """
+
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        glyph_channels=1,
+        position_channels=1,
+    ):
+        super().__init__()
+
+        self.glyph_block = nn.Sequential(
+            nn.Conv2d(glyph_channels, 8, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(8, 8, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(8, 16, 3, padding=1, stride=2),
+            nn.SiLU(),
+            nn.Conv2d(16, 16, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(16, 32, 3, padding=1, stride=2),
+            nn.SiLU(),
+            nn.Conv2d(32, 32, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(32, 96, 3, padding=1, stride=2),
+            nn.SiLU(),
+            nn.Conv2d(96, 96, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(96, 256, 3, padding=1, stride=2),
+            nn.SiLU(),
+        )
+
+        self.position_block = nn.Sequential(
+            nn.Conv2d(position_channels, 8, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(8, 8, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(8, 16, 3, padding=1, stride=2),
+            nn.SiLU(),
+            nn.Conv2d(16, 16, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(16, 32, 3, padding=1, stride=2),
+            nn.SiLU(),
+            nn.Conv2d(32, 32, 3, padding=1),
+            nn.SiLU(),
+            nn.Conv2d(32, 64, 3, padding=1, stride=2),
+            nn.SiLU(),
+        )
+
+        self.fuse_block = nn.Conv2d(256 + 64 + 4, conditioning_embedding_channels, 3, padding=1)
+
+    def forward(self, glyphs, positions, text_info):
+        glyph_embedding = self.glyph_block(glyphs.to(self.glyph_block[0].weight.device))
+        position_embedding = self.position_block(positions.to(self.position_block[0].weight.device))
+        guided_hint = self.fuse_block(torch.cat([glyph_embedding, position_embedding, text_info["masked_x"]], dim=1))
+
+        return guided_hint
+
+
+class AnyTextControlNetModel(ControlNetModel):
+    """
+    A AnyTextControlNetModel model.
+
+    Args:
+        in_channels (`int`, defaults to 4):
+            The number of channels in the input sample.
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, defaults to 0):
+            The frequency shift to apply to the time embedding.
+        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, defaults to 2):
+            The number of layers per block.
+        downsample_padding (`int`, defaults to 1):
+            The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, defaults to 1):
+            The scale factor to use for the mid block.
+        act_fn (`str`, defaults to "silu"):
+            The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+            in post-processing.
+        norm_eps (`float`, defaults to 1e-5):
+            The epsilon to use for the normalization.
+        cross_attention_dim (`int`, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+            The dimension of the attention heads.
+        use_linear_projection (`bool`, defaults to `False`):
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        num_class_embeds (`int`, *optional*, defaults to 0):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        upcast_attention (`bool`, defaults to `False`):
+        resnet_time_scale_shift (`str`, defaults to `"default"`):
+            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+        projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+            The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+            `class_embed_type="projection"`.
+        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+            The tuple of output channel for each block in the `conditioning_embedding` layer.
+        global_pool_conditions (`bool`, defaults to `False`):
+            TODO(Patrick) - unused parameter.
+        addition_embed_type_num_heads (`int`, defaults to 64):
+            The number of heads to use for the `TextTimeEmbedding` layer.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 4,
+        conditioning_channels: int = 1,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str, ...] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+        global_pool_conditions: bool = False,
+        addition_embed_type_num_heads: int = 64,
+    ):
+        super().__init__(
+            in_channels,
+            conditioning_channels,
+            flip_sin_to_cos,
+            freq_shift,
+            down_block_types,
+            mid_block_type,
+            only_cross_attention,
+            block_out_channels,
+            layers_per_block,
+            downsample_padding,
+            mid_block_scale_factor,
+            act_fn,
+            norm_num_groups,
+            norm_eps,
+            cross_attention_dim,
+            transformer_layers_per_block,
+            encoder_hid_dim,
+            encoder_hid_dim_type,
+            attention_head_dim,
+            num_attention_heads,
+            use_linear_projection,
+            class_embed_type,
+            addition_embed_type,
+            addition_time_embed_dim,
+            num_class_embeds,
+            upcast_attention,
+            resnet_time_scale_shift,
+            projection_class_embeddings_input_dim,
+            controlnet_conditioning_channel_order,
+            conditioning_embedding_out_channels,
+            global_pool_conditions,
+            addition_embed_type_num_heads,
+        )
+
+        # control net conditioning embedding
+        self.controlnet_cond_embedding = AnyTextControlNetConditioningEmbedding(
+            conditioning_embedding_channels=block_out_channels[0],
+            glyph_channels=conditioning_channels,
+            position_channels=conditioning_channels,
+        )
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: torch.Tensor,
+        conditioning_scale: float = 1.0,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guess_mode: bool = False,
+        return_dict: bool = True,
+    ) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
+        """
+        The [`~PromptDiffusionControlNetModel`] forward method.
+
+        Args:
+            sample (`torch.Tensor`):
+                The noisy input tensor.
+            timestep (`Union[torch.Tensor, float, int]`):
+                The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states.
+            #controlnet_cond (`torch.Tensor`):
+            #    The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+                embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            added_cond_kwargs (`dict`):
+                Additional conditions for the Stable Diffusion XL UNet.
+            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+            guess_mode (`bool`, defaults to `False`):
+                In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
+                you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+            return_dict (`bool`, defaults to `True`):
+                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
+                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
+                returned where the first element is the sample tensor.
+        """
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order == "rgb":
+            # in rgb order by default
+            ...
+        # elif channel_order == "bgr":
+        #    controlnet_cond = torch.flip(controlnet_cond, dims=[1])
+        else:
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        if self.config.addition_embed_type is not None:
+            if self.config.addition_embed_type == "text":
+                aug_emb = self.add_embedding(encoder_hidden_states)
+
+            elif self.config.addition_embed_type == "text_time":
+                if "text_embeds" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                    )
+                text_embeds = added_cond_kwargs.get("text_embeds")
+                if "time_ids" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                    )
+                time_ids = added_cond_kwargs.get("time_ids")
+                time_embeds = self.add_time_proj(time_ids.flatten())
+                time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+                add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+                add_embeds = add_embeds.to(emb.dtype)
+                aug_emb = self.add_embedding(add_embeds)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        controlnet_cond = self.controlnet_cond_embedding(*controlnet_cond)
+        sample = sample + controlnet_cond
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+        # 5. Control net blocks
+        controlnet_down_block_res_samples = ()
+
+        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+            down_block_res_sample = controlnet_block(down_block_res_sample)
+            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
+
+        down_block_res_samples = controlnet_down_block_res_samples
+
+        mid_block_res_sample = self.controlnet_mid_block(sample)
+
+        # 6. scaling
+        if guess_mode and not self.config.global_pool_conditions:
+            scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
+            scales = scales * conditioning_scale
+            down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
+            mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one
+        else:
+            down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
+            mid_block_res_sample = mid_block_res_sample * conditioning_scale
+
+        if self.config.global_pool_conditions:
+            down_block_res_samples = [
+                torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
+            ]
+            mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)
+
+        if not return_dict:
+            return (down_block_res_samples, mid_block_res_sample)
+
+        return ControlNetOutput(
+            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+        )
+
+
+# Copied from diffusers.models.controlnet.zero_module
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module

examples/research_projects/anytext/ocr_recog/RNN.py

@@ -0,0 +1,209 @@
+import torch
+from torch import nn
+
+from .RecSVTR import Block
+
+
+class Swish(nn.Module):
+    def __int__(self):
+        super(Swish, self).__int__()
+
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class Im2Im(nn.Module):
+    def __init__(self, in_channels, **kwargs):
+        super().__init__()
+        self.out_channels = in_channels
+
+    def forward(self, x):
+        return x
+
+
+class Im2Seq(nn.Module):
+    def __init__(self, in_channels, **kwargs):
+        super().__init__()
+        self.out_channels = in_channels
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        # assert H == 1
+        x = x.reshape(B, C, H * W)
+        x = x.permute((0, 2, 1))
+        return x
+
+
+class EncoderWithRNN(nn.Module):
+    def __init__(self, in_channels, **kwargs):
+        super(EncoderWithRNN, self).__init__()
+        hidden_size = kwargs.get("hidden_size", 256)
+        self.out_channels = hidden_size * 2
+        self.lstm = nn.LSTM(in_channels, hidden_size, bidirectional=True, num_layers=2, batch_first=True)
+
+    def forward(self, x):
+        self.lstm.flatten_parameters()
+        x, _ = self.lstm(x)
+        return x
+
+
+class SequenceEncoder(nn.Module):
+    def __init__(self, in_channels, encoder_type="rnn", **kwargs):
+        super(SequenceEncoder, self).__init__()
+        self.encoder_reshape = Im2Seq(in_channels)
+        self.out_channels = self.encoder_reshape.out_channels
+        self.encoder_type = encoder_type
+        if encoder_type == "reshape":
+            self.only_reshape = True
+        else:
+            support_encoder_dict = {"reshape": Im2Seq, "rnn": EncoderWithRNN, "svtr": EncoderWithSVTR}
+            assert encoder_type in support_encoder_dict, "{} must in {}".format(
+                encoder_type, support_encoder_dict.keys()
+            )
+
+            self.encoder = support_encoder_dict[encoder_type](self.encoder_reshape.out_channels, **kwargs)
+            self.out_channels = self.encoder.out_channels
+            self.only_reshape = False
+
+    def forward(self, x):
+        if self.encoder_type != "svtr":
+            x = self.encoder_reshape(x)
+            if not self.only_reshape:
+                x = self.encoder(x)
+            return x
+        else:
+            x = self.encoder(x)
+            x = self.encoder_reshape(x)
+            return x
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size=3, stride=1, padding=0, bias_attr=False, groups=1, act=nn.GELU
+    ):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            # weight_attr=paddle.ParamAttr(initializer=nn.initializer.KaimingUniform()),
+            bias=bias_attr,
+        )
+        self.norm = nn.BatchNorm2d(out_channels)
+        self.act = Swish()
+
+    def forward(self, inputs):
+        out = self.conv(inputs)
+        out = self.norm(out)
+        out = self.act(out)
+        return out
+
+
+class EncoderWithSVTR(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        dims=64,  # XS
+        depth=2,
+        hidden_dims=120,
+        use_guide=False,
+        num_heads=8,
+        qkv_bias=True,
+        mlp_ratio=2.0,
+        drop_rate=0.1,
+        attn_drop_rate=0.1,
+        drop_path=0.0,
+        qk_scale=None,
+    ):
+        super(EncoderWithSVTR, self).__init__()
+        self.depth = depth
+        self.use_guide = use_guide
+        self.conv1 = ConvBNLayer(in_channels, in_channels // 8, padding=1, act="swish")
+        self.conv2 = ConvBNLayer(in_channels // 8, hidden_dims, kernel_size=1, act="swish")
+
+        self.svtr_block = nn.ModuleList(
+            [
+                Block(
+                    dim=hidden_dims,
+                    num_heads=num_heads,
+                    mixer="Global",
+                    HW=None,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer="swish",
+                    attn_drop=attn_drop_rate,
+                    drop_path=drop_path,
+                    norm_layer="nn.LayerNorm",
+                    epsilon=1e-05,
+                    prenorm=False,
+                )
+                for i in range(depth)
+            ]
+        )
+        self.norm = nn.LayerNorm(hidden_dims, eps=1e-6)
+        self.conv3 = ConvBNLayer(hidden_dims, in_channels, kernel_size=1, act="swish")
+        # last conv-nxn, the input is concat of input tensor and conv3 output tensor
+        self.conv4 = ConvBNLayer(2 * in_channels, in_channels // 8, padding=1, act="swish")
+
+        self.conv1x1 = ConvBNLayer(in_channels // 8, dims, kernel_size=1, act="swish")
+        self.out_channels = dims
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        # weight initialization
+        if isinstance(m, nn.Conv2d):
+            nn.init.kaiming_normal_(m.weight, mode="fan_out")
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.BatchNorm2d):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, 0, 0.01)
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.ConvTranspose2d):
+            nn.init.kaiming_normal_(m.weight, mode="fan_out")
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        # for use guide
+        if self.use_guide:
+            z = x.clone()
+            z.stop_gradient = True
+        else:
+            z = x
+        # for short cut
+        h = z
+        # reduce dim
+        z = self.conv1(z)
+        z = self.conv2(z)
+        # SVTR global block
+        B, C, H, W = z.shape
+        z = z.flatten(2).permute(0, 2, 1)
+
+        for blk in self.svtr_block:
+            z = blk(z)
+
+        z = self.norm(z)
+        # last stage
+        z = z.reshape([-1, H, W, C]).permute(0, 3, 1, 2)
+        z = self.conv3(z)
+        z = torch.cat((h, z), dim=1)
+        z = self.conv1x1(self.conv4(z))
+
+        return z
+
+
+if __name__ == "__main__":
+    svtrRNN = EncoderWithSVTR(56)
+    print(svtrRNN)

examples/research_projects/anytext/ocr_recog/RecCTCHead.py

@@ -0,0 +1,45 @@
+from torch import nn
+
+
+class CTCHead(nn.Module):
+    def __init__(
+        self, in_channels, out_channels=6625, fc_decay=0.0004, mid_channels=None, return_feats=False, **kwargs
+    ):
+        super(CTCHead, self).__init__()
+        if mid_channels is None:
+            self.fc = nn.Linear(
+                in_channels,
+                out_channels,
+                bias=True,
+            )
+        else:
+            self.fc1 = nn.Linear(
+                in_channels,
+                mid_channels,
+                bias=True,
+            )
+            self.fc2 = nn.Linear(
+                mid_channels,
+                out_channels,
+                bias=True,
+            )
+
+        self.out_channels = out_channels
+        self.mid_channels = mid_channels
+        self.return_feats = return_feats
+
+    def forward(self, x, labels=None):
+        if self.mid_channels is None:
+            predicts = self.fc(x)
+        else:
+            x = self.fc1(x)
+            predicts = self.fc2(x)
+
+        if self.return_feats:
+            result = {}
+            result["ctc"] = predicts
+            result["ctc_neck"] = x
+        else:
+            result = predicts
+
+        return result

examples/research_projects/anytext/ocr_recog/RecModel.py

@@ -0,0 +1,49 @@
+from torch import nn
+
+from .RecCTCHead import CTCHead
+from .RecMv1_enhance import MobileNetV1Enhance
+from .RNN import Im2Im, Im2Seq, SequenceEncoder
+
+
+backbone_dict = {"MobileNetV1Enhance": MobileNetV1Enhance}
+neck_dict = {"SequenceEncoder": SequenceEncoder, "Im2Seq": Im2Seq, "None": Im2Im}
+head_dict = {"CTCHead": CTCHead}
+
+
+class RecModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        assert "in_channels" in config, "in_channels must in model config"
+        backbone_type = config["backbone"].pop("type")
+        assert backbone_type in backbone_dict, f"backbone.type must in {backbone_dict}"
+        self.backbone = backbone_dict[backbone_type](config["in_channels"], **config["backbone"])
+
+        neck_type = config["neck"].pop("type")
+        assert neck_type in neck_dict, f"neck.type must in {neck_dict}"
+        self.neck = neck_dict[neck_type](self.backbone.out_channels, **config["neck"])
+
+        head_type = config["head"].pop("type")
+        assert head_type in head_dict, f"head.type must in {head_dict}"
+        self.head = head_dict[head_type](self.neck.out_channels, **config["head"])
+
+        self.name = f"RecModel_{backbone_type}_{neck_type}_{head_type}"
+
+    def load_3rd_state_dict(self, _3rd_name, _state):
+        self.backbone.load_3rd_state_dict(_3rd_name, _state)
+        self.neck.load_3rd_state_dict(_3rd_name, _state)
+        self.head.load_3rd_state_dict(_3rd_name, _state)
+
+    def forward(self, x):
+        import torch
+
+        x = x.to(torch.float32)
+        x = self.backbone(x)
+        x = self.neck(x)
+        x = self.head(x)
+        return x
+
+    def encode(self, x):
+        x = self.backbone(x)
+        x = self.neck(x)
+        x = self.head.ctc_encoder(x)
+        return x

examples/research_projects/anytext/ocr_recog/RecMv1_enhance.py

@@ -0,0 +1,197 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from .common import Activation
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self, num_channels, filter_size, num_filters, stride, padding, channels=None, num_groups=1, act="hard_swish"
+    ):
+        super(ConvBNLayer, self).__init__()
+        self.act = act
+        self._conv = nn.Conv2d(
+            in_channels=num_channels,
+            out_channels=num_filters,
+            kernel_size=filter_size,
+            stride=stride,
+            padding=padding,
+            groups=num_groups,
+            bias=False,
+        )
+
+        self._batch_norm = nn.BatchNorm2d(
+            num_filters,
+        )
+        if self.act is not None:
+            self._act = Activation(act_type=act, inplace=True)
+
+    def forward(self, inputs):
+        y = self._conv(inputs)
+        y = self._batch_norm(y)
+        if self.act is not None:
+            y = self._act(y)
+        return y
+
+
+class DepthwiseSeparable(nn.Module):
+    def __init__(
+        self, num_channels, num_filters1, num_filters2, num_groups, stride, scale, dw_size=3, padding=1, use_se=False
+    ):
+        super(DepthwiseSeparable, self).__init__()
+        self.use_se = use_se
+        self._depthwise_conv = ConvBNLayer(
+            num_channels=num_channels,
+            num_filters=int(num_filters1 * scale),
+            filter_size=dw_size,
+            stride=stride,
+            padding=padding,
+            num_groups=int(num_groups * scale),
+        )
+        if use_se:
+            self._se = SEModule(int(num_filters1 * scale))
+        self._pointwise_conv = ConvBNLayer(
+            num_channels=int(num_filters1 * scale),
+            filter_size=1,
+            num_filters=int(num_filters2 * scale),
+            stride=1,
+            padding=0,
+        )
+
+    def forward(self, inputs):
+        y = self._depthwise_conv(inputs)
+        if self.use_se:
+            y = self._se(y)
+        y = self._pointwise_conv(y)
+        return y
+
+
+class MobileNetV1Enhance(nn.Module):
+    def __init__(self, in_channels=3, scale=0.5, last_conv_stride=1, last_pool_type="max", **kwargs):
+        super().__init__()
+        self.scale = scale
+        self.block_list = []
+
+        self.conv1 = ConvBNLayer(
+            num_channels=in_channels, filter_size=3, channels=3, num_filters=int(32 * scale), stride=2, padding=1
+        )
+
+        conv2_1 = DepthwiseSeparable(
+            num_channels=int(32 * scale), num_filters1=32, num_filters2=64, num_groups=32, stride=1, scale=scale
+        )
+        self.block_list.append(conv2_1)
+
+        conv2_2 = DepthwiseSeparable(
+            num_channels=int(64 * scale), num_filters1=64, num_filters2=128, num_groups=64, stride=1, scale=scale
+        )
+        self.block_list.append(conv2_2)
+
+        conv3_1 = DepthwiseSeparable(
+            num_channels=int(128 * scale), num_filters1=128, num_filters2=128, num_groups=128, stride=1, scale=scale
+        )
+        self.block_list.append(conv3_1)
+
+        conv3_2 = DepthwiseSeparable(
+            num_channels=int(128 * scale),
+            num_filters1=128,
+            num_filters2=256,
+            num_groups=128,
+            stride=(2, 1),
+            scale=scale,
+        )
+        self.block_list.append(conv3_2)
+
+        conv4_1 = DepthwiseSeparable(
+            num_channels=int(256 * scale), num_filters1=256, num_filters2=256, num_groups=256, stride=1, scale=scale
+        )
+        self.block_list.append(conv4_1)
+
+        conv4_2 = DepthwiseSeparable(
+            num_channels=int(256 * scale),
+            num_filters1=256,
+            num_filters2=512,
+            num_groups=256,
+            stride=(2, 1),
+            scale=scale,
+        )
+        self.block_list.append(conv4_2)
+
+        for _ in range(5):
+            conv5 = DepthwiseSeparable(
+                num_channels=int(512 * scale),
+                num_filters1=512,
+                num_filters2=512,
+                num_groups=512,
+                stride=1,
+                dw_size=5,
+                padding=2,
+                scale=scale,
+                use_se=False,
+            )
+            self.block_list.append(conv5)
+
+        conv5_6 = DepthwiseSeparable(
+            num_channels=int(512 * scale),
+            num_filters1=512,
+            num_filters2=1024,
+            num_groups=512,
+            stride=(2, 1),
+            dw_size=5,
+            padding=2,
+            scale=scale,
+            use_se=True,
+        )
+        self.block_list.append(conv5_6)
+
+        conv6 = DepthwiseSeparable(
+            num_channels=int(1024 * scale),
+            num_filters1=1024,
+            num_filters2=1024,
+            num_groups=1024,
+            stride=last_conv_stride,
+            dw_size=5,
+            padding=2,
+            use_se=True,
+            scale=scale,
+        )
+        self.block_list.append(conv6)
+
+        self.block_list = nn.Sequential(*self.block_list)
+        if last_pool_type == "avg":
+            self.pool = nn.AvgPool2d(kernel_size=2, stride=2, padding=0)
+        else:
+            self.pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
+        self.out_channels = int(1024 * scale)
+
+    def forward(self, inputs):
+        y = self.conv1(inputs)
+        y = self.block_list(y)
+        y = self.pool(y)
+        return y
+
+
+def hardsigmoid(x):
+    return F.relu6(x + 3.0, inplace=True) / 6.0
+
+
+class SEModule(nn.Module):
+    def __init__(self, channel, reduction=4):
+        super(SEModule, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.conv1 = nn.Conv2d(
+            in_channels=channel, out_channels=channel // reduction, kernel_size=1, stride=1, padding=0, bias=True
+        )
+        self.conv2 = nn.Conv2d(
+            in_channels=channel // reduction, out_channels=channel, kernel_size=1, stride=1, padding=0, bias=True
+        )
+
+    def forward(self, inputs):
+        outputs = self.avg_pool(inputs)
+        outputs = self.conv1(outputs)
+        outputs = F.relu(outputs)
+        outputs = self.conv2(outputs)
+        outputs = hardsigmoid(outputs)
+        x = torch.mul(inputs, outputs)
+
+        return x

examples/research_projects/anytext/ocr_recog/RecSVTR.py

@@ -0,0 +1,570 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from torch.nn import functional
+from torch.nn.init import ones_, trunc_normal_, zeros_
+
+
+def drop_path(x, drop_prob=0.0, training=False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ...
+    """
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = torch.tensor(1 - drop_prob)
+    shape = (x.size()[0],) + (1,) * (x.ndim - 1)
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype)
+    random_tensor = torch.floor(random_tensor)  # binarize
+    output = x.divide(keep_prob) * random_tensor
+    return output
+
+
+class Swish(nn.Module):
+    def __int__(self):
+        super(Swish, self).__int__()
+
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size=3, stride=1, padding=0, bias_attr=False, groups=1, act=nn.GELU
+    ):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            # weight_attr=paddle.ParamAttr(initializer=nn.initializer.KaimingUniform()),
+            bias=bias_attr,
+        )
+        self.norm = nn.BatchNorm2d(out_channels)
+        self.act = act()
+
+    def forward(self, inputs):
+        out = self.conv(inputs)
+        out = self.norm(out)
+        out = self.act(out)
+        return out
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class Identity(nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, input):
+        return input
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.0):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        if isinstance(act_layer, str):
+            self.act = Swish()
+        else:
+            self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class ConvMixer(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        HW=(8, 25),
+        local_k=(3, 3),
+    ):
+        super().__init__()
+        self.HW = HW
+        self.dim = dim
+        self.local_mixer = nn.Conv2d(
+            dim,
+            dim,
+            local_k,
+            1,
+            (local_k[0] // 2, local_k[1] // 2),
+            groups=num_heads,
+            # weight_attr=ParamAttr(initializer=KaimingNormal())
+        )
+
+    def forward(self, x):
+        h = self.HW[0]
+        w = self.HW[1]
+        x = x.transpose([0, 2, 1]).reshape([0, self.dim, h, w])
+        x = self.local_mixer(x)
+        x = x.flatten(2).transpose([0, 2, 1])
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        mixer="Global",
+        HW=(8, 25),
+        local_k=(7, 11),
+        qkv_bias=False,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.HW = HW
+        if HW is not None:
+            H = HW[0]
+            W = HW[1]
+            self.N = H * W
+            self.C = dim
+        if mixer == "Local" and HW is not None:
+            hk = local_k[0]
+            wk = local_k[1]
+            mask = torch.ones([H * W, H + hk - 1, W + wk - 1])
+            for h in range(0, H):
+                for w in range(0, W):
+                    mask[h * W + w, h : h + hk, w : w + wk] = 0.0
+            mask_paddle = mask[:, hk // 2 : H + hk // 2, wk // 2 : W + wk // 2].flatten(1)
+            mask_inf = torch.full([H * W, H * W], fill_value=float("-inf"))
+            mask = torch.where(mask_paddle < 1, mask_paddle, mask_inf)
+            self.mask = mask[None, None, :]
+            # self.mask = mask.unsqueeze([0, 1])
+        self.mixer = mixer
+
+    def forward(self, x):
+        if self.HW is not None:
+            N = self.N
+            C = self.C
+        else:
+            _, N, C = x.shape
+        qkv = self.qkv(x).reshape((-1, N, 3, self.num_heads, C // self.num_heads)).permute((2, 0, 3, 1, 4))
+        q, k, v = qkv[0] * self.scale, qkv[1], qkv[2]
+
+        attn = q.matmul(k.permute((0, 1, 3, 2)))
+        if self.mixer == "Local":
+            attn += self.mask
+        attn = functional.softmax(attn, dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn.matmul(v)).permute((0, 2, 1, 3)).reshape((-1, N, C))
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        mixer="Global",
+        local_mixer=(7, 11),
+        HW=(8, 25),
+        mlp_ratio=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer=nn.GELU,
+        norm_layer="nn.LayerNorm",
+        epsilon=1e-6,
+        prenorm=True,
+    ):
+        super().__init__()
+        if isinstance(norm_layer, str):
+            self.norm1 = eval(norm_layer)(dim, eps=epsilon)
+        else:
+            self.norm1 = norm_layer(dim)
+        if mixer == "Global" or mixer == "Local":
+            self.mixer = Attention(
+                dim,
+                num_heads=num_heads,
+                mixer=mixer,
+                HW=HW,
+                local_k=local_mixer,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                attn_drop=attn_drop,
+                proj_drop=drop,
+            )
+        elif mixer == "Conv":
+            self.mixer = ConvMixer(dim, num_heads=num_heads, HW=HW, local_k=local_mixer)
+        else:
+            raise TypeError("The mixer must be one of [Global, Local, Conv]")
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else Identity()
+        if isinstance(norm_layer, str):
+            self.norm2 = eval(norm_layer)(dim, eps=epsilon)
+        else:
+            self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp_ratio = mlp_ratio
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+        self.prenorm = prenorm
+
+    def forward(self, x):
+        if self.prenorm:
+            x = self.norm1(x + self.drop_path(self.mixer(x)))
+            x = self.norm2(x + self.drop_path(self.mlp(x)))
+        else:
+            x = x + self.drop_path(self.mixer(self.norm1(x)))
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding"""
+
+    def __init__(self, img_size=(32, 100), in_channels=3, embed_dim=768, sub_num=2):
+        super().__init__()
+        num_patches = (img_size[1] // (2**sub_num)) * (img_size[0] // (2**sub_num))
+        self.img_size = img_size
+        self.num_patches = num_patches
+        self.embed_dim = embed_dim
+        self.norm = None
+        if sub_num == 2:
+            self.proj = nn.Sequential(
+                ConvBNLayer(
+                    in_channels=in_channels,
+                    out_channels=embed_dim // 2,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    act=nn.GELU,
+                    bias_attr=False,
+                ),
+                ConvBNLayer(
+                    in_channels=embed_dim // 2,
+                    out_channels=embed_dim,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    act=nn.GELU,
+                    bias_attr=False,
+                ),
+            )
+        if sub_num == 3:
+            self.proj = nn.Sequential(
+                ConvBNLayer(
+                    in_channels=in_channels,
+                    out_channels=embed_dim // 4,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    act=nn.GELU,
+                    bias_attr=False,
+                ),
+                ConvBNLayer(
+                    in_channels=embed_dim // 4,
+                    out_channels=embed_dim // 2,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    act=nn.GELU,
+                    bias_attr=False,
+                ),
+                ConvBNLayer(
+                    in_channels=embed_dim // 2,
+                    out_channels=embed_dim,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    act=nn.GELU,
+                    bias_attr=False,
+                ),
+            )
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        assert (
+            H == self.img_size[0] and W == self.img_size[1]
+        ), f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).permute(0, 2, 1)
+        return x
+
+
+class SubSample(nn.Module):
+    def __init__(self, in_channels, out_channels, types="Pool", stride=(2, 1), sub_norm="nn.LayerNorm", act=None):
+        super().__init__()
+        self.types = types
+        if types == "Pool":
+            self.avgpool = nn.AvgPool2d(kernel_size=(3, 5), stride=stride, padding=(1, 2))
+            self.maxpool = nn.MaxPool2d(kernel_size=(3, 5), stride=stride, padding=(1, 2))
+            self.proj = nn.Linear(in_channels, out_channels)
+        else:
+            self.conv = nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=3,
+                stride=stride,
+                padding=1,
+                # weight_attr=ParamAttr(initializer=KaimingNormal())
+            )
+        self.norm = eval(sub_norm)(out_channels)
+        if act is not None:
+            self.act = act()
+        else:
+            self.act = None
+
+    def forward(self, x):
+        if self.types == "Pool":
+            x1 = self.avgpool(x)
+            x2 = self.maxpool(x)
+            x = (x1 + x2) * 0.5
+            out = self.proj(x.flatten(2).permute((0, 2, 1)))
+        else:
+            x = self.conv(x)
+            out = x.flatten(2).permute((0, 2, 1))
+        out = self.norm(out)
+        if self.act is not None:
+            out = self.act(out)
+
+        return out
+
+
+class SVTRNet(nn.Module):
+    def __init__(
+        self,
+        img_size=[48, 100],
+        in_channels=3,
+        embed_dim=[64, 128, 256],
+        depth=[3, 6, 3],
+        num_heads=[2, 4, 8],
+        mixer=["Local"] * 6 + ["Global"] * 6,  # Local atten, Global atten, Conv
+        local_mixer=[[7, 11], [7, 11], [7, 11]],
+        patch_merging="Conv",  # Conv, Pool, None
+        mlp_ratio=4,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.0,
+        last_drop=0.1,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.1,
+        norm_layer="nn.LayerNorm",
+        sub_norm="nn.LayerNorm",
+        epsilon=1e-6,
+        out_channels=192,
+        out_char_num=25,
+        block_unit="Block",
+        act="nn.GELU",
+        last_stage=True,
+        sub_num=2,
+        prenorm=True,
+        use_lenhead=False,
+        **kwargs,
+    ):
+        super().__init__()
+        self.img_size = img_size
+        self.embed_dim = embed_dim
+        self.out_channels = out_channels
+        self.prenorm = prenorm
+        patch_merging = None if patch_merging != "Conv" and patch_merging != "Pool" else patch_merging
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, in_channels=in_channels, embed_dim=embed_dim[0], sub_num=sub_num
+        )
+        num_patches = self.patch_embed.num_patches
+        self.HW = [img_size[0] // (2**sub_num), img_size[1] // (2**sub_num)]
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim[0]))
+        # self.pos_embed = self.create_parameter(
+        #     shape=[1, num_patches, embed_dim[0]], default_initializer=zeros_)
+
+        # self.add_parameter("pos_embed", self.pos_embed)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+        Block_unit = eval(block_unit)
+
+        dpr = np.linspace(0, drop_path_rate, sum(depth))
+        self.blocks1 = nn.ModuleList(
+            [
+                Block_unit(
+                    dim=embed_dim[0],
+                    num_heads=num_heads[0],
+                    mixer=mixer[0 : depth[0]][i],
+                    HW=self.HW,
+                    local_mixer=local_mixer[0],
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer=eval(act),
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[0 : depth[0]][i],
+                    norm_layer=norm_layer,
+                    epsilon=epsilon,
+                    prenorm=prenorm,
+                )
+                for i in range(depth[0])
+            ]
+        )
+        if patch_merging is not None:
+            self.sub_sample1 = SubSample(
+                embed_dim[0], embed_dim[1], sub_norm=sub_norm, stride=[2, 1], types=patch_merging
+            )
+            HW = [self.HW[0] // 2, self.HW[1]]
+        else:
+            HW = self.HW
+        self.patch_merging = patch_merging
+        self.blocks2 = nn.ModuleList(
+            [
+                Block_unit(
+                    dim=embed_dim[1],
+                    num_heads=num_heads[1],
+                    mixer=mixer[depth[0] : depth[0] + depth[1]][i],
+                    HW=HW,
+                    local_mixer=local_mixer[1],
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer=eval(act),
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[depth[0] : depth[0] + depth[1]][i],
+                    norm_layer=norm_layer,
+                    epsilon=epsilon,
+                    prenorm=prenorm,
+                )
+                for i in range(depth[1])
+            ]
+        )
+        if patch_merging is not None:
+            self.sub_sample2 = SubSample(
+                embed_dim[1], embed_dim[2], sub_norm=sub_norm, stride=[2, 1], types=patch_merging
+            )
+            HW = [self.HW[0] // 4, self.HW[1]]
+        else:
+            HW = self.HW
+        self.blocks3 = nn.ModuleList(
+            [
+                Block_unit(
+                    dim=embed_dim[2],
+                    num_heads=num_heads[2],
+                    mixer=mixer[depth[0] + depth[1] :][i],
+                    HW=HW,
+                    local_mixer=local_mixer[2],
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer=eval(act),
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[depth[0] + depth[1] :][i],
+                    norm_layer=norm_layer,
+                    epsilon=epsilon,
+                    prenorm=prenorm,
+                )
+                for i in range(depth[2])
+            ]
+        )
+        self.last_stage = last_stage
+        if last_stage:
+            self.avg_pool = nn.AdaptiveAvgPool2d((1, out_char_num))
+            self.last_conv = nn.Conv2d(
+                in_channels=embed_dim[2],
+                out_channels=self.out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=False,
+            )
+            self.hardswish = nn.Hardswish()
+            self.dropout = nn.Dropout(p=last_drop)
+        if not prenorm:
+            self.norm = eval(norm_layer)(embed_dim[-1], epsilon=epsilon)
+        self.use_lenhead = use_lenhead
+        if use_lenhead:
+            self.len_conv = nn.Linear(embed_dim[2], self.out_channels)
+            self.hardswish_len = nn.Hardswish()
+            self.dropout_len = nn.Dropout(p=last_drop)
+
+        trunc_normal_(self.pos_embed, std=0.02)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                zeros_(m.bias)
+        elif isinstance(m, nn.LayerNorm):
+            zeros_(m.bias)
+            ones_(m.weight)
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        x = x + self.pos_embed
+        x = self.pos_drop(x)
+        for blk in self.blocks1:
+            x = blk(x)
+        if self.patch_merging is not None:
+            x = self.sub_sample1(x.permute([0, 2, 1]).reshape([-1, self.embed_dim[0], self.HW[0], self.HW[1]]))
+        for blk in self.blocks2:
+            x = blk(x)
+        if self.patch_merging is not None:
+            x = self.sub_sample2(x.permute([0, 2, 1]).reshape([-1, self.embed_dim[1], self.HW[0] // 2, self.HW[1]]))
+        for blk in self.blocks3:
+            x = blk(x)
+        if not self.prenorm:
+            x = self.norm(x)
+        return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        if self.use_lenhead:
+            len_x = self.len_conv(x.mean(1))
+            len_x = self.dropout_len(self.hardswish_len(len_x))
+        if self.last_stage:
+            if self.patch_merging is not None:
+                h = self.HW[0] // 4
+            else:
+                h = self.HW[0]
+            x = self.avg_pool(x.permute([0, 2, 1]).reshape([-1, self.embed_dim[2], h, self.HW[1]]))
+            x = self.last_conv(x)
+            x = self.hardswish(x)
+            x = self.dropout(x)
+        if self.use_lenhead:
+            return x, len_x
+        return x
+
+
+if __name__ == "__main__":
+    a = torch.rand(1, 3, 48, 100)
+    svtr = SVTRNet()
+
+    out = svtr(a)
+    print(svtr)
+    print(out.size())

examples/research_projects/anytext/ocr_recog/common.py

@@ -0,0 +1,74 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Hswish(nn.Module):
+    def __init__(self, inplace=True):
+        super(Hswish, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0
+
+
+# out = max(0, min(1, slop*x+offset))
+# paddle.fluid.layers.hard_sigmoid(x, slope=0.2, offset=0.5, name=None)
+class Hsigmoid(nn.Module):
+    def __init__(self, inplace=True):
+        super(Hsigmoid, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        # torch: F.relu6(x + 3., inplace=self.inplace) / 6.
+        # paddle: F.relu6(1.2 * x + 3., inplace=self.inplace) / 6.
+        return F.relu6(1.2 * x + 3.0, inplace=self.inplace) / 6.0
+
+
+class GELU(nn.Module):
+    def __init__(self, inplace=True):
+        super(GELU, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        return torch.nn.functional.gelu(x)
+
+
+class Swish(nn.Module):
+    def __init__(self, inplace=True):
+        super(Swish, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        if self.inplace:
+            x.mul_(torch.sigmoid(x))
+            return x
+        else:
+            return x * torch.sigmoid(x)
+
+
+class Activation(nn.Module):
+    def __init__(self, act_type, inplace=True):
+        super(Activation, self).__init__()
+        act_type = act_type.lower()
+        if act_type == "relu":
+            self.act = nn.ReLU(inplace=inplace)
+        elif act_type == "relu6":
+            self.act = nn.ReLU6(inplace=inplace)
+        elif act_type == "sigmoid":
+            raise NotImplementedError
+        elif act_type == "hard_sigmoid":
+            self.act = Hsigmoid(inplace)
+        elif act_type == "hard_swish":
+            self.act = Hswish(inplace=inplace)
+        elif act_type == "leakyrelu":
+            self.act = nn.LeakyReLU(inplace=inplace)
+        elif act_type == "gelu":
+            self.act = GELU(inplace=inplace)
+        elif act_type == "swish":
+            self.act = Swish(inplace=inplace)
+        else:
+            raise NotImplementedError
+
+    def forward(self, inputs):
+        return self.act(inputs)

examples/research_projects/anytext/ocr_recog/en_dict.txt

@@ -0,0 +1,95 @@
+0
+1
+2
+3
+4
+5
+6
+7
+8
+9
+:
+;
+<
+=
+>
+?
+@
+A
+B
+C
+D
+E
+F
+G
+H
+I
+J
+K
+L
+M
+N
+O
+P
+Q
+R
+S
+T
+U
+V
+W
+X
+Y
+Z
+[
+\
+]
+^
+_
+`
+a
+b
+c
+d
+e
+f
+g
+h
+i
+j
+k
+l
+m
+n
+o
+p
+q
+r
+s
+t
+u
+v
+w
+x
+y
+z
+{
+|
+}
+~
+!
+"
+#
+$
+%
+&
+'
+(
+)
+*
++
+,
+-
+.
+/
+ 

examples/research_projects/realfill/requirements.txt

@@ -6,4 +6,4 @@ torch==2.2.0
 torchvision>=0.16
 ftfy==6.1.1
 tensorboard==2.14.0
-Jinja2==3.1.5
+Jinja2==3.1.6